open-nomad/scheduler/reconcile_util.go

Ignoring revisions in .git-blame-ignore-revs. Click here to bypass and see the normal blame view.

822 lines
25 KiB
Go
Raw Permalink Normal View History

// Copyright (c) HashiCorp, Inc.
// SPDX-License-Identifier: MPL-2.0
2017-05-23 00:14:38 +00:00
package scheduler
// The structs and helpers in this file are split out of reconciler.go for code
// manageability and should not be shared to the system schedulers! If you need
// something here for system/sysbatch jobs, double-check it's safe to use for
// all scheduler types before moving it into util.go
2017-05-23 00:14:38 +00:00
import (
"errors"
2017-05-23 00:14:38 +00:00
"fmt"
2017-06-02 23:11:29 +00:00
"sort"
"strings"
"time"
2017-05-23 00:14:38 +00:00
"github.com/hashicorp/nomad/nomad/structs"
)
2017-12-13 17:36:03 +00:00
// placementResult is an allocation that must be placed. It potentially has a
2017-07-15 23:31:33 +00:00
// previous allocation attached to it that should be stopped only if the
// paired placement is complete. This gives an atomic place/stop behavior to
// prevent an impossible resource ask as part of a rolling update to wipe the
// job out.
type placementResult interface {
// TaskGroup returns the task group the placement is for
TaskGroup() *structs.TaskGroup
// Name returns the name of the desired allocation
Name() string
// Canary returns whether the placement should be a canary
Canary() bool
// PreviousAllocation returns the previous allocation
PreviousAllocation() *structs.Allocation
2018-01-19 21:20:00 +00:00
// IsRescheduling returns whether the placement was rescheduling a failed allocation
IsRescheduling() bool
2017-07-15 23:31:33 +00:00
// StopPreviousAlloc returns whether the previous allocation should be
// stopped and if so the status description.
StopPreviousAlloc() (bool, string)
// PreviousLost is true if the previous allocation was lost.
PreviousLost() bool
// DowngradeNonCanary indicates that placement should use the latest stable job
// with the MinJobVersion, rather than the current deployment version
DowngradeNonCanary() bool
MinJobVersion() uint64
2017-07-15 23:31:33 +00:00
}
// allocStopResult contains the information required to stop a single allocation
type allocStopResult struct {
alloc *structs.Allocation
clientStatus string
statusDescription string
followupEvalID string
2017-07-15 23:31:33 +00:00
}
// allocPlaceResult contains the information required to place a single
// allocation
type allocPlaceResult struct {
name string
canary bool
taskGroup *structs.TaskGroup
previousAlloc *structs.Allocation
reschedule bool
lost bool
downgradeNonCanary bool
minJobVersion uint64
2017-07-15 23:31:33 +00:00
}
func (a allocPlaceResult) TaskGroup() *structs.TaskGroup { return a.taskGroup }
func (a allocPlaceResult) Name() string { return a.name }
func (a allocPlaceResult) Canary() bool { return a.canary }
func (a allocPlaceResult) PreviousAllocation() *structs.Allocation { return a.previousAlloc }
2018-01-19 21:20:00 +00:00
func (a allocPlaceResult) IsRescheduling() bool { return a.reschedule }
2017-07-15 23:31:33 +00:00
func (a allocPlaceResult) StopPreviousAlloc() (bool, string) { return false, "" }
func (a allocPlaceResult) DowngradeNonCanary() bool { return a.downgradeNonCanary }
func (a allocPlaceResult) MinJobVersion() uint64 { return a.minJobVersion }
func (a allocPlaceResult) PreviousLost() bool { return a.lost }
2017-07-15 23:31:33 +00:00
// allocDestructiveResult contains the information required to do a destructive
// update. Destructive changes should be applied atomically, as in the old alloc
// is only stopped if the new one can be placed.
type allocDestructiveResult struct {
placeName string
placeTaskGroup *structs.TaskGroup
stopAlloc *structs.Allocation
stopStatusDescription string
}
func (a allocDestructiveResult) TaskGroup() *structs.TaskGroup { return a.placeTaskGroup }
func (a allocDestructiveResult) Name() string { return a.placeName }
func (a allocDestructiveResult) Canary() bool { return false }
func (a allocDestructiveResult) PreviousAllocation() *structs.Allocation { return a.stopAlloc }
2018-01-19 21:20:00 +00:00
func (a allocDestructiveResult) IsRescheduling() bool { return false }
2017-07-15 23:31:33 +00:00
func (a allocDestructiveResult) StopPreviousAlloc() (bool, string) {
return true, a.stopStatusDescription
}
func (a allocDestructiveResult) DowngradeNonCanary() bool { return false }
func (a allocDestructiveResult) MinJobVersion() uint64 { return 0 }
func (a allocDestructiveResult) PreviousLost() bool { return false }
2017-07-15 23:31:33 +00:00
2017-05-23 00:14:38 +00:00
// allocMatrix is a mapping of task groups to their allocation set.
type allocMatrix map[string]allocSet
// newAllocMatrix takes a job and the existing allocations for the job and
// creates an allocMatrix
func newAllocMatrix(job *structs.Job, allocs []*structs.Allocation) allocMatrix {
m := allocMatrix(make(map[string]allocSet))
for _, a := range allocs {
s, ok := m[a.TaskGroup]
if !ok {
s = make(map[string]*structs.Allocation)
m[a.TaskGroup] = s
}
s[a.ID] = a
}
2017-06-01 22:16:24 +00:00
if job != nil {
for _, tg := range job.TaskGroups {
2017-09-26 22:26:33 +00:00
if _, ok := m[tg.Name]; !ok {
m[tg.Name] = make(map[string]*structs.Allocation)
2017-06-01 22:16:24 +00:00
}
2017-05-23 00:14:38 +00:00
}
}
return m
}
// allocSet is a set of allocations with a series of helper functions defined
// that help reconcile state.
type allocSet map[string]*structs.Allocation
// GoString provides a human readable view of the set
func (a allocSet) GoString() string {
if len(a) == 0 {
return "[]"
}
start := fmt.Sprintf("len(%d) [\n", len(a))
2017-06-02 23:11:29 +00:00
var s []string
for k, v := range a {
s = append(s, fmt.Sprintf("%q: %v", k, v.Name))
}
return start + strings.Join(s, "\n") + "]"
}
2017-06-06 21:08:46 +00:00
// nameSet returns the set of allocation names
2017-06-02 23:11:29 +00:00
func (a allocSet) nameSet() map[string]struct{} {
names := make(map[string]struct{}, len(a))
for _, alloc := range a {
names[alloc.Name] = struct{}{}
}
return names
}
2017-06-06 21:08:46 +00:00
// nameOrder returns the set of allocation names in sorted order
2017-06-02 23:11:29 +00:00
func (a allocSet) nameOrder() []*structs.Allocation {
allocs := make([]*structs.Allocation, 0, len(a))
for _, alloc := range a {
allocs = append(allocs, alloc)
2017-05-23 00:14:38 +00:00
}
2017-06-02 23:11:29 +00:00
sort.Slice(allocs, func(i, j int) bool {
return allocs[i].Index() < allocs[j].Index()
})
return allocs
2017-05-23 00:14:38 +00:00
}
// difference returns a new allocSet that has all the existing item except those
// contained within the other allocation sets
func (a allocSet) difference(others ...allocSet) allocSet {
diff := make(map[string]*structs.Allocation)
OUTER:
for k, v := range a {
for _, other := range others {
if _, ok := other[k]; ok {
continue OUTER
}
}
diff[k] = v
}
return diff
}
2017-05-31 18:34:46 +00:00
// union returns a new allocSet that has the union of the two allocSets.
// Conflicts prefer the last passed allocSet containing the value
func (a allocSet) union(others ...allocSet) allocSet {
union := make(map[string]*structs.Allocation, len(a))
order := []allocSet{a}
order = append(order, others...)
for _, set := range order {
for k, v := range set {
union[k] = v
}
}
return union
}
2017-07-05 19:50:40 +00:00
// fromKeys returns an alloc set matching the passed keys
func (a allocSet) fromKeys(keys ...[]string) allocSet {
from := make(map[string]*structs.Allocation)
for _, set := range keys {
for _, k := range set {
if alloc, ok := a[k]; ok {
from[k] = alloc
}
}
}
return from
}
2017-12-13 17:36:03 +00:00
// filterByTainted takes a set of tainted nodes and filters the allocation set
// into the following groups:
2017-05-23 00:14:38 +00:00
// 1. Those that exist on untainted nodes
// 2. Those exist on nodes that are draining
// 3. Those that exist on lost nodes or have expired
// 4. Those that are on nodes that are disconnected, but have not had their ClientState set to unknown
// 5. Those that are on a node that has reconnected.
// 6. Those that are in a state that results in a noop.
func (a allocSet) filterByTainted(taintedNodes map[string]*structs.Node, serverSupportsDisconnectedClients bool, now time.Time) (untainted, migrate, lost, disconnecting, reconnecting, ignore allocSet) {
2017-05-23 00:14:38 +00:00
untainted = make(map[string]*structs.Allocation)
migrate = make(map[string]*structs.Allocation)
lost = make(map[string]*structs.Allocation)
disconnecting = make(map[string]*structs.Allocation)
reconnecting = make(map[string]*structs.Allocation)
ignore = make(map[string]*structs.Allocation)
2017-05-23 00:14:38 +00:00
for _, alloc := range a {
// make sure we don't apply any reconnect logic to task groups
// without max_client_disconnect
supportsDisconnectedClients := alloc.SupportsDisconnectedClients(serverSupportsDisconnectedClients)
Update alloc after reconnect and enforece client heartbeat order (#15068) * scheduler: allow updates after alloc reconnects When an allocation reconnects to a cluster the scheduler needs to run special logic to handle the reconnection, check if a replacement was create and stop one of them. If the allocation kept running while the node was disconnected, it will be reconnected with `ClientStatus: running` and the node will have `Status: ready`. This combination is the same as the normal steady state of allocation, where everything is running as expected. In order to differentiate between the two states (an allocation that is reconnecting and one that is just running) the scheduler needs an extra piece of state. The current implementation uses the presence of a `TaskClientReconnected` task event to detect when the allocation has reconnected and thus must go through the reconnection process. But this event remains even after the allocation is reconnected, causing all future evals to consider the allocation as still reconnecting. This commit changes the reconnect logic to use an `AllocState` to register when the allocation was reconnected. This provides the following benefits: - Only a limited number of task states are kept, and they are used for many other events. It's possible that, upon reconnecting, several actions are triggered that could cause the `TaskClientReconnected` event to be dropped. - Task events are set by clients and so their timestamps are subject to time skew from servers. This prevents using time to determine if an allocation reconnected after a disconnect event. - Disconnect events are already stored as `AllocState` and so storing reconnects there as well makes it the only source of information required. With the new logic, the reconnection logic is only triggered if the last `AllocState` is a disconnect event, meaning that the allocation has not been reconnected yet. After the reconnection is handled, the new `ClientStatus` is store in `AllocState` allowing future evals to skip the reconnection logic. * scheduler: prevent spurious placement on reconnect When a client reconnects it makes two independent RPC calls: - `Node.UpdateStatus` to heartbeat and set its status as `ready`. - `Node.UpdateAlloc` to update the status of its allocations. These two calls can happen in any order, and in case the allocations are updated before a heartbeat it causes the state to be the same as a node being disconnected: the node status will still be `disconnected` while the allocation `ClientStatus` is set to `running`. The current implementation did not handle this order of events properly, and the scheduler would create an unnecessary placement since it considered the allocation was being disconnected. This extra allocation would then be quickly stopped by the heartbeat eval. This commit adds a new code path to handle this order of events. If the node is `disconnected` and the allocation `ClientStatus` is `running` the scheduler will check if the allocation is actually reconnecting using its `AllocState` events. * rpc: only allow alloc updates from `ready` nodes Clients interact with servers using three main RPC methods: - `Node.GetAllocs` reads allocation data from the server and writes it to the client. - `Node.UpdateAlloc` reads allocation from from the client and writes them to the server. - `Node.UpdateStatus` writes the client status to the server and is used as the heartbeat mechanism. These three methods are called periodically by the clients and are done so independently from each other, meaning that there can't be any assumptions in their ordering. This can generate scenarios that are hard to reason about and to code for. For example, when a client misses too many heartbeats it will be considered `down` or `disconnected` and the allocations it was running are set to `lost` or `unknown`. When connectivity is restored the to rest of the cluster, the natural mental model is to think that the client will heartbeat first and then update its allocations status into the servers. But since there's no inherit order in these calls the reverse is just as possible: the client updates the alloc status and then heartbeats. This results in a state where allocs are, for example, `running` while the client is still `disconnected`. This commit adds a new verification to the `Node.UpdateAlloc` method to reject updates from nodes that are not `ready`, forcing clients to heartbeat first. Since this check is done server-side there is no need to coordinate operations client-side: they can continue sending these requests independently and alloc update will succeed after the heartbeat is done. * chagelog: add entry for #15068 * code review * client: skip terminal allocations on reconnect When the client reconnects with the server it synchronizes the state of its allocations by sending data using the `Node.UpdateAlloc` RPC and fetching data using the `Node.GetClientAllocs` RPC. If the data fetch happens before the data write, `unknown` allocations will still be in this state and would trigger the `allocRunner.Reconnect` flow. But when the server `DesiredStatus` for the allocation is `stop` the client should not reconnect the allocation. * apply more code review changes * scheduler: persist changes to reconnected allocs Reconnected allocs have a new AllocState entry that must be persisted by the plan applier. * rpc: read node ID from allocs in UpdateAlloc The AllocUpdateRequest struct is used in three disjoint use cases: 1. Stripped allocs from clients Node.UpdateAlloc RPC using the Allocs, and WriteRequest fields 2. Raft log message using the Allocs, Evals, and WriteRequest fields 3. Plan updates using the AllocsStopped, AllocsUpdated, and Job fields Adding a new field that would only be used in one these cases (1) made things more confusing and error prone. While in theory an AllocUpdateRequest could send allocations from different nodes, in practice this never actually happens since only clients call this method with their own allocations. * scheduler: remove logic to handle exceptional case This condition could only be hit if, somehow, the allocation status was set to "running" while the client was "unknown". This was addressed by enforcing an order in "Node.UpdateStatus" and "Node.UpdateAlloc" RPC calls, so this scenario is not expected to happen. Adding unnecessary code to the scheduler makes it harder to read and reason about it. * more code review * remove another unused test
2022-11-04 20:25:11 +00:00
reconnect := false
expired := false
Update alloc after reconnect and enforece client heartbeat order (#15068) * scheduler: allow updates after alloc reconnects When an allocation reconnects to a cluster the scheduler needs to run special logic to handle the reconnection, check if a replacement was create and stop one of them. If the allocation kept running while the node was disconnected, it will be reconnected with `ClientStatus: running` and the node will have `Status: ready`. This combination is the same as the normal steady state of allocation, where everything is running as expected. In order to differentiate between the two states (an allocation that is reconnecting and one that is just running) the scheduler needs an extra piece of state. The current implementation uses the presence of a `TaskClientReconnected` task event to detect when the allocation has reconnected and thus must go through the reconnection process. But this event remains even after the allocation is reconnected, causing all future evals to consider the allocation as still reconnecting. This commit changes the reconnect logic to use an `AllocState` to register when the allocation was reconnected. This provides the following benefits: - Only a limited number of task states are kept, and they are used for many other events. It's possible that, upon reconnecting, several actions are triggered that could cause the `TaskClientReconnected` event to be dropped. - Task events are set by clients and so their timestamps are subject to time skew from servers. This prevents using time to determine if an allocation reconnected after a disconnect event. - Disconnect events are already stored as `AllocState` and so storing reconnects there as well makes it the only source of information required. With the new logic, the reconnection logic is only triggered if the last `AllocState` is a disconnect event, meaning that the allocation has not been reconnected yet. After the reconnection is handled, the new `ClientStatus` is store in `AllocState` allowing future evals to skip the reconnection logic. * scheduler: prevent spurious placement on reconnect When a client reconnects it makes two independent RPC calls: - `Node.UpdateStatus` to heartbeat and set its status as `ready`. - `Node.UpdateAlloc` to update the status of its allocations. These two calls can happen in any order, and in case the allocations are updated before a heartbeat it causes the state to be the same as a node being disconnected: the node status will still be `disconnected` while the allocation `ClientStatus` is set to `running`. The current implementation did not handle this order of events properly, and the scheduler would create an unnecessary placement since it considered the allocation was being disconnected. This extra allocation would then be quickly stopped by the heartbeat eval. This commit adds a new code path to handle this order of events. If the node is `disconnected` and the allocation `ClientStatus` is `running` the scheduler will check if the allocation is actually reconnecting using its `AllocState` events. * rpc: only allow alloc updates from `ready` nodes Clients interact with servers using three main RPC methods: - `Node.GetAllocs` reads allocation data from the server and writes it to the client. - `Node.UpdateAlloc` reads allocation from from the client and writes them to the server. - `Node.UpdateStatus` writes the client status to the server and is used as the heartbeat mechanism. These three methods are called periodically by the clients and are done so independently from each other, meaning that there can't be any assumptions in their ordering. This can generate scenarios that are hard to reason about and to code for. For example, when a client misses too many heartbeats it will be considered `down` or `disconnected` and the allocations it was running are set to `lost` or `unknown`. When connectivity is restored the to rest of the cluster, the natural mental model is to think that the client will heartbeat first and then update its allocations status into the servers. But since there's no inherit order in these calls the reverse is just as possible: the client updates the alloc status and then heartbeats. This results in a state where allocs are, for example, `running` while the client is still `disconnected`. This commit adds a new verification to the `Node.UpdateAlloc` method to reject updates from nodes that are not `ready`, forcing clients to heartbeat first. Since this check is done server-side there is no need to coordinate operations client-side: they can continue sending these requests independently and alloc update will succeed after the heartbeat is done. * chagelog: add entry for #15068 * code review * client: skip terminal allocations on reconnect When the client reconnects with the server it synchronizes the state of its allocations by sending data using the `Node.UpdateAlloc` RPC and fetching data using the `Node.GetClientAllocs` RPC. If the data fetch happens before the data write, `unknown` allocations will still be in this state and would trigger the `allocRunner.Reconnect` flow. But when the server `DesiredStatus` for the allocation is `stop` the client should not reconnect the allocation. * apply more code review changes * scheduler: persist changes to reconnected allocs Reconnected allocs have a new AllocState entry that must be persisted by the plan applier. * rpc: read node ID from allocs in UpdateAlloc The AllocUpdateRequest struct is used in three disjoint use cases: 1. Stripped allocs from clients Node.UpdateAlloc RPC using the Allocs, and WriteRequest fields 2. Raft log message using the Allocs, Evals, and WriteRequest fields 3. Plan updates using the AllocsStopped, AllocsUpdated, and Job fields Adding a new field that would only be used in one these cases (1) made things more confusing and error prone. While in theory an AllocUpdateRequest could send allocations from different nodes, in practice this never actually happens since only clients call this method with their own allocations. * scheduler: remove logic to handle exceptional case This condition could only be hit if, somehow, the allocation status was set to "running" while the client was "unknown". This was addressed by enforcing an order in "Node.UpdateStatus" and "Node.UpdateAlloc" RPC calls, so this scenario is not expected to happen. Adding unnecessary code to the scheduler makes it harder to read and reason about it. * more code review * remove another unused test
2022-11-04 20:25:11 +00:00
// Only compute reconnect for unknown, running, and failed since they
// need to go through the reconnect logic.
if supportsDisconnectedClients &&
(alloc.ClientStatus == structs.AllocClientStatusUnknown ||
alloc.ClientStatus == structs.AllocClientStatusRunning ||
alloc.ClientStatus == structs.AllocClientStatusFailed) {
Update alloc after reconnect and enforece client heartbeat order (#15068) * scheduler: allow updates after alloc reconnects When an allocation reconnects to a cluster the scheduler needs to run special logic to handle the reconnection, check if a replacement was create and stop one of them. If the allocation kept running while the node was disconnected, it will be reconnected with `ClientStatus: running` and the node will have `Status: ready`. This combination is the same as the normal steady state of allocation, where everything is running as expected. In order to differentiate between the two states (an allocation that is reconnecting and one that is just running) the scheduler needs an extra piece of state. The current implementation uses the presence of a `TaskClientReconnected` task event to detect when the allocation has reconnected and thus must go through the reconnection process. But this event remains even after the allocation is reconnected, causing all future evals to consider the allocation as still reconnecting. This commit changes the reconnect logic to use an `AllocState` to register when the allocation was reconnected. This provides the following benefits: - Only a limited number of task states are kept, and they are used for many other events. It's possible that, upon reconnecting, several actions are triggered that could cause the `TaskClientReconnected` event to be dropped. - Task events are set by clients and so their timestamps are subject to time skew from servers. This prevents using time to determine if an allocation reconnected after a disconnect event. - Disconnect events are already stored as `AllocState` and so storing reconnects there as well makes it the only source of information required. With the new logic, the reconnection logic is only triggered if the last `AllocState` is a disconnect event, meaning that the allocation has not been reconnected yet. After the reconnection is handled, the new `ClientStatus` is store in `AllocState` allowing future evals to skip the reconnection logic. * scheduler: prevent spurious placement on reconnect When a client reconnects it makes two independent RPC calls: - `Node.UpdateStatus` to heartbeat and set its status as `ready`. - `Node.UpdateAlloc` to update the status of its allocations. These two calls can happen in any order, and in case the allocations are updated before a heartbeat it causes the state to be the same as a node being disconnected: the node status will still be `disconnected` while the allocation `ClientStatus` is set to `running`. The current implementation did not handle this order of events properly, and the scheduler would create an unnecessary placement since it considered the allocation was being disconnected. This extra allocation would then be quickly stopped by the heartbeat eval. This commit adds a new code path to handle this order of events. If the node is `disconnected` and the allocation `ClientStatus` is `running` the scheduler will check if the allocation is actually reconnecting using its `AllocState` events. * rpc: only allow alloc updates from `ready` nodes Clients interact with servers using three main RPC methods: - `Node.GetAllocs` reads allocation data from the server and writes it to the client. - `Node.UpdateAlloc` reads allocation from from the client and writes them to the server. - `Node.UpdateStatus` writes the client status to the server and is used as the heartbeat mechanism. These three methods are called periodically by the clients and are done so independently from each other, meaning that there can't be any assumptions in their ordering. This can generate scenarios that are hard to reason about and to code for. For example, when a client misses too many heartbeats it will be considered `down` or `disconnected` and the allocations it was running are set to `lost` or `unknown`. When connectivity is restored the to rest of the cluster, the natural mental model is to think that the client will heartbeat first and then update its allocations status into the servers. But since there's no inherit order in these calls the reverse is just as possible: the client updates the alloc status and then heartbeats. This results in a state where allocs are, for example, `running` while the client is still `disconnected`. This commit adds a new verification to the `Node.UpdateAlloc` method to reject updates from nodes that are not `ready`, forcing clients to heartbeat first. Since this check is done server-side there is no need to coordinate operations client-side: they can continue sending these requests independently and alloc update will succeed after the heartbeat is done. * chagelog: add entry for #15068 * code review * client: skip terminal allocations on reconnect When the client reconnects with the server it synchronizes the state of its allocations by sending data using the `Node.UpdateAlloc` RPC and fetching data using the `Node.GetClientAllocs` RPC. If the data fetch happens before the data write, `unknown` allocations will still be in this state and would trigger the `allocRunner.Reconnect` flow. But when the server `DesiredStatus` for the allocation is `stop` the client should not reconnect the allocation. * apply more code review changes * scheduler: persist changes to reconnected allocs Reconnected allocs have a new AllocState entry that must be persisted by the plan applier. * rpc: read node ID from allocs in UpdateAlloc The AllocUpdateRequest struct is used in three disjoint use cases: 1. Stripped allocs from clients Node.UpdateAlloc RPC using the Allocs, and WriteRequest fields 2. Raft log message using the Allocs, Evals, and WriteRequest fields 3. Plan updates using the AllocsStopped, AllocsUpdated, and Job fields Adding a new field that would only be used in one these cases (1) made things more confusing and error prone. While in theory an AllocUpdateRequest could send allocations from different nodes, in practice this never actually happens since only clients call this method with their own allocations. * scheduler: remove logic to handle exceptional case This condition could only be hit if, somehow, the allocation status was set to "running" while the client was "unknown". This was addressed by enforcing an order in "Node.UpdateStatus" and "Node.UpdateAlloc" RPC calls, so this scenario is not expected to happen. Adding unnecessary code to the scheduler makes it harder to read and reason about it. * more code review * remove another unused test
2022-11-04 20:25:11 +00:00
reconnect = alloc.NeedsToReconnect()
if reconnect {
expired = alloc.Expired(now)
}
}
Update alloc after reconnect and enforece client heartbeat order (#15068) * scheduler: allow updates after alloc reconnects When an allocation reconnects to a cluster the scheduler needs to run special logic to handle the reconnection, check if a replacement was create and stop one of them. If the allocation kept running while the node was disconnected, it will be reconnected with `ClientStatus: running` and the node will have `Status: ready`. This combination is the same as the normal steady state of allocation, where everything is running as expected. In order to differentiate between the two states (an allocation that is reconnecting and one that is just running) the scheduler needs an extra piece of state. The current implementation uses the presence of a `TaskClientReconnected` task event to detect when the allocation has reconnected and thus must go through the reconnection process. But this event remains even after the allocation is reconnected, causing all future evals to consider the allocation as still reconnecting. This commit changes the reconnect logic to use an `AllocState` to register when the allocation was reconnected. This provides the following benefits: - Only a limited number of task states are kept, and they are used for many other events. It's possible that, upon reconnecting, several actions are triggered that could cause the `TaskClientReconnected` event to be dropped. - Task events are set by clients and so their timestamps are subject to time skew from servers. This prevents using time to determine if an allocation reconnected after a disconnect event. - Disconnect events are already stored as `AllocState` and so storing reconnects there as well makes it the only source of information required. With the new logic, the reconnection logic is only triggered if the last `AllocState` is a disconnect event, meaning that the allocation has not been reconnected yet. After the reconnection is handled, the new `ClientStatus` is store in `AllocState` allowing future evals to skip the reconnection logic. * scheduler: prevent spurious placement on reconnect When a client reconnects it makes two independent RPC calls: - `Node.UpdateStatus` to heartbeat and set its status as `ready`. - `Node.UpdateAlloc` to update the status of its allocations. These two calls can happen in any order, and in case the allocations are updated before a heartbeat it causes the state to be the same as a node being disconnected: the node status will still be `disconnected` while the allocation `ClientStatus` is set to `running`. The current implementation did not handle this order of events properly, and the scheduler would create an unnecessary placement since it considered the allocation was being disconnected. This extra allocation would then be quickly stopped by the heartbeat eval. This commit adds a new code path to handle this order of events. If the node is `disconnected` and the allocation `ClientStatus` is `running` the scheduler will check if the allocation is actually reconnecting using its `AllocState` events. * rpc: only allow alloc updates from `ready` nodes Clients interact with servers using three main RPC methods: - `Node.GetAllocs` reads allocation data from the server and writes it to the client. - `Node.UpdateAlloc` reads allocation from from the client and writes them to the server. - `Node.UpdateStatus` writes the client status to the server and is used as the heartbeat mechanism. These three methods are called periodically by the clients and are done so independently from each other, meaning that there can't be any assumptions in their ordering. This can generate scenarios that are hard to reason about and to code for. For example, when a client misses too many heartbeats it will be considered `down` or `disconnected` and the allocations it was running are set to `lost` or `unknown`. When connectivity is restored the to rest of the cluster, the natural mental model is to think that the client will heartbeat first and then update its allocations status into the servers. But since there's no inherit order in these calls the reverse is just as possible: the client updates the alloc status and then heartbeats. This results in a state where allocs are, for example, `running` while the client is still `disconnected`. This commit adds a new verification to the `Node.UpdateAlloc` method to reject updates from nodes that are not `ready`, forcing clients to heartbeat first. Since this check is done server-side there is no need to coordinate operations client-side: they can continue sending these requests independently and alloc update will succeed after the heartbeat is done. * chagelog: add entry for #15068 * code review * client: skip terminal allocations on reconnect When the client reconnects with the server it synchronizes the state of its allocations by sending data using the `Node.UpdateAlloc` RPC and fetching data using the `Node.GetClientAllocs` RPC. If the data fetch happens before the data write, `unknown` allocations will still be in this state and would trigger the `allocRunner.Reconnect` flow. But when the server `DesiredStatus` for the allocation is `stop` the client should not reconnect the allocation. * apply more code review changes * scheduler: persist changes to reconnected allocs Reconnected allocs have a new AllocState entry that must be persisted by the plan applier. * rpc: read node ID from allocs in UpdateAlloc The AllocUpdateRequest struct is used in three disjoint use cases: 1. Stripped allocs from clients Node.UpdateAlloc RPC using the Allocs, and WriteRequest fields 2. Raft log message using the Allocs, Evals, and WriteRequest fields 3. Plan updates using the AllocsStopped, AllocsUpdated, and Job fields Adding a new field that would only be used in one these cases (1) made things more confusing and error prone. While in theory an AllocUpdateRequest could send allocations from different nodes, in practice this never actually happens since only clients call this method with their own allocations. * scheduler: remove logic to handle exceptional case This condition could only be hit if, somehow, the allocation status was set to "running" while the client was "unknown". This was addressed by enforcing an order in "Node.UpdateStatus" and "Node.UpdateAlloc" RPC calls, so this scenario is not expected to happen. Adding unnecessary code to the scheduler makes it harder to read and reason about it. * more code review * remove another unused test
2022-11-04 20:25:11 +00:00
// Failed allocs that need to be reconnected must be added to
// reconnecting so that they can be handled as a failed reconnect.
if supportsDisconnectedClients &&
Update alloc after reconnect and enforece client heartbeat order (#15068) * scheduler: allow updates after alloc reconnects When an allocation reconnects to a cluster the scheduler needs to run special logic to handle the reconnection, check if a replacement was create and stop one of them. If the allocation kept running while the node was disconnected, it will be reconnected with `ClientStatus: running` and the node will have `Status: ready`. This combination is the same as the normal steady state of allocation, where everything is running as expected. In order to differentiate between the two states (an allocation that is reconnecting and one that is just running) the scheduler needs an extra piece of state. The current implementation uses the presence of a `TaskClientReconnected` task event to detect when the allocation has reconnected and thus must go through the reconnection process. But this event remains even after the allocation is reconnected, causing all future evals to consider the allocation as still reconnecting. This commit changes the reconnect logic to use an `AllocState` to register when the allocation was reconnected. This provides the following benefits: - Only a limited number of task states are kept, and they are used for many other events. It's possible that, upon reconnecting, several actions are triggered that could cause the `TaskClientReconnected` event to be dropped. - Task events are set by clients and so their timestamps are subject to time skew from servers. This prevents using time to determine if an allocation reconnected after a disconnect event. - Disconnect events are already stored as `AllocState` and so storing reconnects there as well makes it the only source of information required. With the new logic, the reconnection logic is only triggered if the last `AllocState` is a disconnect event, meaning that the allocation has not been reconnected yet. After the reconnection is handled, the new `ClientStatus` is store in `AllocState` allowing future evals to skip the reconnection logic. * scheduler: prevent spurious placement on reconnect When a client reconnects it makes two independent RPC calls: - `Node.UpdateStatus` to heartbeat and set its status as `ready`. - `Node.UpdateAlloc` to update the status of its allocations. These two calls can happen in any order, and in case the allocations are updated before a heartbeat it causes the state to be the same as a node being disconnected: the node status will still be `disconnected` while the allocation `ClientStatus` is set to `running`. The current implementation did not handle this order of events properly, and the scheduler would create an unnecessary placement since it considered the allocation was being disconnected. This extra allocation would then be quickly stopped by the heartbeat eval. This commit adds a new code path to handle this order of events. If the node is `disconnected` and the allocation `ClientStatus` is `running` the scheduler will check if the allocation is actually reconnecting using its `AllocState` events. * rpc: only allow alloc updates from `ready` nodes Clients interact with servers using three main RPC methods: - `Node.GetAllocs` reads allocation data from the server and writes it to the client. - `Node.UpdateAlloc` reads allocation from from the client and writes them to the server. - `Node.UpdateStatus` writes the client status to the server and is used as the heartbeat mechanism. These three methods are called periodically by the clients and are done so independently from each other, meaning that there can't be any assumptions in their ordering. This can generate scenarios that are hard to reason about and to code for. For example, when a client misses too many heartbeats it will be considered `down` or `disconnected` and the allocations it was running are set to `lost` or `unknown`. When connectivity is restored the to rest of the cluster, the natural mental model is to think that the client will heartbeat first and then update its allocations status into the servers. But since there's no inherit order in these calls the reverse is just as possible: the client updates the alloc status and then heartbeats. This results in a state where allocs are, for example, `running` while the client is still `disconnected`. This commit adds a new verification to the `Node.UpdateAlloc` method to reject updates from nodes that are not `ready`, forcing clients to heartbeat first. Since this check is done server-side there is no need to coordinate operations client-side: they can continue sending these requests independently and alloc update will succeed after the heartbeat is done. * chagelog: add entry for #15068 * code review * client: skip terminal allocations on reconnect When the client reconnects with the server it synchronizes the state of its allocations by sending data using the `Node.UpdateAlloc` RPC and fetching data using the `Node.GetClientAllocs` RPC. If the data fetch happens before the data write, `unknown` allocations will still be in this state and would trigger the `allocRunner.Reconnect` flow. But when the server `DesiredStatus` for the allocation is `stop` the client should not reconnect the allocation. * apply more code review changes * scheduler: persist changes to reconnected allocs Reconnected allocs have a new AllocState entry that must be persisted by the plan applier. * rpc: read node ID from allocs in UpdateAlloc The AllocUpdateRequest struct is used in three disjoint use cases: 1. Stripped allocs from clients Node.UpdateAlloc RPC using the Allocs, and WriteRequest fields 2. Raft log message using the Allocs, Evals, and WriteRequest fields 3. Plan updates using the AllocsStopped, AllocsUpdated, and Job fields Adding a new field that would only be used in one these cases (1) made things more confusing and error prone. While in theory an AllocUpdateRequest could send allocations from different nodes, in practice this never actually happens since only clients call this method with their own allocations. * scheduler: remove logic to handle exceptional case This condition could only be hit if, somehow, the allocation status was set to "running" while the client was "unknown". This was addressed by enforcing an order in "Node.UpdateStatus" and "Node.UpdateAlloc" RPC calls, so this scenario is not expected to happen. Adding unnecessary code to the scheduler makes it harder to read and reason about it. * more code review * remove another unused test
2022-11-04 20:25:11 +00:00
reconnect &&
alloc.DesiredStatus == structs.AllocDesiredStatusRun &&
alloc.ClientStatus == structs.AllocClientStatusFailed {
reconnecting[alloc.ID] = alloc
continue
}
taintedNode, nodeIsTainted := taintedNodes[alloc.NodeID]
if taintedNode != nil {
// Group disconnecting
switch taintedNode.Status {
case structs.NodeStatusDisconnected:
if supportsDisconnectedClients {
// Filter running allocs on a node that is disconnected to be marked as unknown.
if alloc.ClientStatus == structs.AllocClientStatusRunning {
disconnecting[alloc.ID] = alloc
continue
}
// Filter pending allocs on a node that is disconnected to be marked as lost.
if alloc.ClientStatus == structs.AllocClientStatusPending {
lost[alloc.ID] = alloc
continue
}
} else {
lost[alloc.ID] = alloc
continue
}
case structs.NodeStatusReady:
// Filter reconnecting allocs on a node that is now connected.
Update alloc after reconnect and enforece client heartbeat order (#15068) * scheduler: allow updates after alloc reconnects When an allocation reconnects to a cluster the scheduler needs to run special logic to handle the reconnection, check if a replacement was create and stop one of them. If the allocation kept running while the node was disconnected, it will be reconnected with `ClientStatus: running` and the node will have `Status: ready`. This combination is the same as the normal steady state of allocation, where everything is running as expected. In order to differentiate between the two states (an allocation that is reconnecting and one that is just running) the scheduler needs an extra piece of state. The current implementation uses the presence of a `TaskClientReconnected` task event to detect when the allocation has reconnected and thus must go through the reconnection process. But this event remains even after the allocation is reconnected, causing all future evals to consider the allocation as still reconnecting. This commit changes the reconnect logic to use an `AllocState` to register when the allocation was reconnected. This provides the following benefits: - Only a limited number of task states are kept, and they are used for many other events. It's possible that, upon reconnecting, several actions are triggered that could cause the `TaskClientReconnected` event to be dropped. - Task events are set by clients and so their timestamps are subject to time skew from servers. This prevents using time to determine if an allocation reconnected after a disconnect event. - Disconnect events are already stored as `AllocState` and so storing reconnects there as well makes it the only source of information required. With the new logic, the reconnection logic is only triggered if the last `AllocState` is a disconnect event, meaning that the allocation has not been reconnected yet. After the reconnection is handled, the new `ClientStatus` is store in `AllocState` allowing future evals to skip the reconnection logic. * scheduler: prevent spurious placement on reconnect When a client reconnects it makes two independent RPC calls: - `Node.UpdateStatus` to heartbeat and set its status as `ready`. - `Node.UpdateAlloc` to update the status of its allocations. These two calls can happen in any order, and in case the allocations are updated before a heartbeat it causes the state to be the same as a node being disconnected: the node status will still be `disconnected` while the allocation `ClientStatus` is set to `running`. The current implementation did not handle this order of events properly, and the scheduler would create an unnecessary placement since it considered the allocation was being disconnected. This extra allocation would then be quickly stopped by the heartbeat eval. This commit adds a new code path to handle this order of events. If the node is `disconnected` and the allocation `ClientStatus` is `running` the scheduler will check if the allocation is actually reconnecting using its `AllocState` events. * rpc: only allow alloc updates from `ready` nodes Clients interact with servers using three main RPC methods: - `Node.GetAllocs` reads allocation data from the server and writes it to the client. - `Node.UpdateAlloc` reads allocation from from the client and writes them to the server. - `Node.UpdateStatus` writes the client status to the server and is used as the heartbeat mechanism. These three methods are called periodically by the clients and are done so independently from each other, meaning that there can't be any assumptions in their ordering. This can generate scenarios that are hard to reason about and to code for. For example, when a client misses too many heartbeats it will be considered `down` or `disconnected` and the allocations it was running are set to `lost` or `unknown`. When connectivity is restored the to rest of the cluster, the natural mental model is to think that the client will heartbeat first and then update its allocations status into the servers. But since there's no inherit order in these calls the reverse is just as possible: the client updates the alloc status and then heartbeats. This results in a state where allocs are, for example, `running` while the client is still `disconnected`. This commit adds a new verification to the `Node.UpdateAlloc` method to reject updates from nodes that are not `ready`, forcing clients to heartbeat first. Since this check is done server-side there is no need to coordinate operations client-side: they can continue sending these requests independently and alloc update will succeed after the heartbeat is done. * chagelog: add entry for #15068 * code review * client: skip terminal allocations on reconnect When the client reconnects with the server it synchronizes the state of its allocations by sending data using the `Node.UpdateAlloc` RPC and fetching data using the `Node.GetClientAllocs` RPC. If the data fetch happens before the data write, `unknown` allocations will still be in this state and would trigger the `allocRunner.Reconnect` flow. But when the server `DesiredStatus` for the allocation is `stop` the client should not reconnect the allocation. * apply more code review changes * scheduler: persist changes to reconnected allocs Reconnected allocs have a new AllocState entry that must be persisted by the plan applier. * rpc: read node ID from allocs in UpdateAlloc The AllocUpdateRequest struct is used in three disjoint use cases: 1. Stripped allocs from clients Node.UpdateAlloc RPC using the Allocs, and WriteRequest fields 2. Raft log message using the Allocs, Evals, and WriteRequest fields 3. Plan updates using the AllocsStopped, AllocsUpdated, and Job fields Adding a new field that would only be used in one these cases (1) made things more confusing and error prone. While in theory an AllocUpdateRequest could send allocations from different nodes, in practice this never actually happens since only clients call this method with their own allocations. * scheduler: remove logic to handle exceptional case This condition could only be hit if, somehow, the allocation status was set to "running" while the client was "unknown". This was addressed by enforcing an order in "Node.UpdateStatus" and "Node.UpdateAlloc" RPC calls, so this scenario is not expected to happen. Adding unnecessary code to the scheduler makes it harder to read and reason about it. * more code review * remove another unused test
2022-11-04 20:25:11 +00:00
if reconnect {
if expired {
lost[alloc.ID] = alloc
continue
}
reconnecting[alloc.ID] = alloc
continue
}
default:
}
}
Update alloc after reconnect and enforece client heartbeat order (#15068) * scheduler: allow updates after alloc reconnects When an allocation reconnects to a cluster the scheduler needs to run special logic to handle the reconnection, check if a replacement was create and stop one of them. If the allocation kept running while the node was disconnected, it will be reconnected with `ClientStatus: running` and the node will have `Status: ready`. This combination is the same as the normal steady state of allocation, where everything is running as expected. In order to differentiate between the two states (an allocation that is reconnecting and one that is just running) the scheduler needs an extra piece of state. The current implementation uses the presence of a `TaskClientReconnected` task event to detect when the allocation has reconnected and thus must go through the reconnection process. But this event remains even after the allocation is reconnected, causing all future evals to consider the allocation as still reconnecting. This commit changes the reconnect logic to use an `AllocState` to register when the allocation was reconnected. This provides the following benefits: - Only a limited number of task states are kept, and they are used for many other events. It's possible that, upon reconnecting, several actions are triggered that could cause the `TaskClientReconnected` event to be dropped. - Task events are set by clients and so their timestamps are subject to time skew from servers. This prevents using time to determine if an allocation reconnected after a disconnect event. - Disconnect events are already stored as `AllocState` and so storing reconnects there as well makes it the only source of information required. With the new logic, the reconnection logic is only triggered if the last `AllocState` is a disconnect event, meaning that the allocation has not been reconnected yet. After the reconnection is handled, the new `ClientStatus` is store in `AllocState` allowing future evals to skip the reconnection logic. * scheduler: prevent spurious placement on reconnect When a client reconnects it makes two independent RPC calls: - `Node.UpdateStatus` to heartbeat and set its status as `ready`. - `Node.UpdateAlloc` to update the status of its allocations. These two calls can happen in any order, and in case the allocations are updated before a heartbeat it causes the state to be the same as a node being disconnected: the node status will still be `disconnected` while the allocation `ClientStatus` is set to `running`. The current implementation did not handle this order of events properly, and the scheduler would create an unnecessary placement since it considered the allocation was being disconnected. This extra allocation would then be quickly stopped by the heartbeat eval. This commit adds a new code path to handle this order of events. If the node is `disconnected` and the allocation `ClientStatus` is `running` the scheduler will check if the allocation is actually reconnecting using its `AllocState` events. * rpc: only allow alloc updates from `ready` nodes Clients interact with servers using three main RPC methods: - `Node.GetAllocs` reads allocation data from the server and writes it to the client. - `Node.UpdateAlloc` reads allocation from from the client and writes them to the server. - `Node.UpdateStatus` writes the client status to the server and is used as the heartbeat mechanism. These three methods are called periodically by the clients and are done so independently from each other, meaning that there can't be any assumptions in their ordering. This can generate scenarios that are hard to reason about and to code for. For example, when a client misses too many heartbeats it will be considered `down` or `disconnected` and the allocations it was running are set to `lost` or `unknown`. When connectivity is restored the to rest of the cluster, the natural mental model is to think that the client will heartbeat first and then update its allocations status into the servers. But since there's no inherit order in these calls the reverse is just as possible: the client updates the alloc status and then heartbeats. This results in a state where allocs are, for example, `running` while the client is still `disconnected`. This commit adds a new verification to the `Node.UpdateAlloc` method to reject updates from nodes that are not `ready`, forcing clients to heartbeat first. Since this check is done server-side there is no need to coordinate operations client-side: they can continue sending these requests independently and alloc update will succeed after the heartbeat is done. * chagelog: add entry for #15068 * code review * client: skip terminal allocations on reconnect When the client reconnects with the server it synchronizes the state of its allocations by sending data using the `Node.UpdateAlloc` RPC and fetching data using the `Node.GetClientAllocs` RPC. If the data fetch happens before the data write, `unknown` allocations will still be in this state and would trigger the `allocRunner.Reconnect` flow. But when the server `DesiredStatus` for the allocation is `stop` the client should not reconnect the allocation. * apply more code review changes * scheduler: persist changes to reconnected allocs Reconnected allocs have a new AllocState entry that must be persisted by the plan applier. * rpc: read node ID from allocs in UpdateAlloc The AllocUpdateRequest struct is used in three disjoint use cases: 1. Stripped allocs from clients Node.UpdateAlloc RPC using the Allocs, and WriteRequest fields 2. Raft log message using the Allocs, Evals, and WriteRequest fields 3. Plan updates using the AllocsStopped, AllocsUpdated, and Job fields Adding a new field that would only be used in one these cases (1) made things more confusing and error prone. While in theory an AllocUpdateRequest could send allocations from different nodes, in practice this never actually happens since only clients call this method with their own allocations. * scheduler: remove logic to handle exceptional case This condition could only be hit if, somehow, the allocation status was set to "running" while the client was "unknown". This was addressed by enforcing an order in "Node.UpdateStatus" and "Node.UpdateAlloc" RPC calls, so this scenario is not expected to happen. Adding unnecessary code to the scheduler makes it harder to read and reason about it. * more code review * remove another unused test
2022-11-04 20:25:11 +00:00
if alloc.TerminalStatus() && !reconnect {
// Terminal allocs, if supportsDisconnectedClient and not reconnect,
// are probably stopped replacements and should be ignored
if supportsDisconnectedClients {
ignore[alloc.ID] = alloc
continue
}
// Terminal allocs, if not reconnect, are always untainted as they
// should never be migrated.
2017-05-24 20:43:01 +00:00
untainted[alloc.ID] = alloc
continue
}
// Non-terminal allocs that should migrate should always migrate
if alloc.DesiredTransition.ShouldMigrate() {
migrate[alloc.ID] = alloc
2017-05-24 20:43:01 +00:00
continue
}
// Expired unknown allocs are lost
if supportsDisconnectedClients && alloc.Expired(now) {
lost[alloc.ID] = alloc
continue
}
// Acknowledge unknown allocs that we want to reconnect eventually.
if supportsDisconnectedClients &&
alloc.ClientStatus == structs.AllocClientStatusUnknown &&
alloc.DesiredStatus == structs.AllocDesiredStatusRun {
untainted[alloc.ID] = alloc
continue
}
Update alloc after reconnect and enforece client heartbeat order (#15068) * scheduler: allow updates after alloc reconnects When an allocation reconnects to a cluster the scheduler needs to run special logic to handle the reconnection, check if a replacement was create and stop one of them. If the allocation kept running while the node was disconnected, it will be reconnected with `ClientStatus: running` and the node will have `Status: ready`. This combination is the same as the normal steady state of allocation, where everything is running as expected. In order to differentiate between the two states (an allocation that is reconnecting and one that is just running) the scheduler needs an extra piece of state. The current implementation uses the presence of a `TaskClientReconnected` task event to detect when the allocation has reconnected and thus must go through the reconnection process. But this event remains even after the allocation is reconnected, causing all future evals to consider the allocation as still reconnecting. This commit changes the reconnect logic to use an `AllocState` to register when the allocation was reconnected. This provides the following benefits: - Only a limited number of task states are kept, and they are used for many other events. It's possible that, upon reconnecting, several actions are triggered that could cause the `TaskClientReconnected` event to be dropped. - Task events are set by clients and so their timestamps are subject to time skew from servers. This prevents using time to determine if an allocation reconnected after a disconnect event. - Disconnect events are already stored as `AllocState` and so storing reconnects there as well makes it the only source of information required. With the new logic, the reconnection logic is only triggered if the last `AllocState` is a disconnect event, meaning that the allocation has not been reconnected yet. After the reconnection is handled, the new `ClientStatus` is store in `AllocState` allowing future evals to skip the reconnection logic. * scheduler: prevent spurious placement on reconnect When a client reconnects it makes two independent RPC calls: - `Node.UpdateStatus` to heartbeat and set its status as `ready`. - `Node.UpdateAlloc` to update the status of its allocations. These two calls can happen in any order, and in case the allocations are updated before a heartbeat it causes the state to be the same as a node being disconnected: the node status will still be `disconnected` while the allocation `ClientStatus` is set to `running`. The current implementation did not handle this order of events properly, and the scheduler would create an unnecessary placement since it considered the allocation was being disconnected. This extra allocation would then be quickly stopped by the heartbeat eval. This commit adds a new code path to handle this order of events. If the node is `disconnected` and the allocation `ClientStatus` is `running` the scheduler will check if the allocation is actually reconnecting using its `AllocState` events. * rpc: only allow alloc updates from `ready` nodes Clients interact with servers using three main RPC methods: - `Node.GetAllocs` reads allocation data from the server and writes it to the client. - `Node.UpdateAlloc` reads allocation from from the client and writes them to the server. - `Node.UpdateStatus` writes the client status to the server and is used as the heartbeat mechanism. These three methods are called periodically by the clients and are done so independently from each other, meaning that there can't be any assumptions in their ordering. This can generate scenarios that are hard to reason about and to code for. For example, when a client misses too many heartbeats it will be considered `down` or `disconnected` and the allocations it was running are set to `lost` or `unknown`. When connectivity is restored the to rest of the cluster, the natural mental model is to think that the client will heartbeat first and then update its allocations status into the servers. But since there's no inherit order in these calls the reverse is just as possible: the client updates the alloc status and then heartbeats. This results in a state where allocs are, for example, `running` while the client is still `disconnected`. This commit adds a new verification to the `Node.UpdateAlloc` method to reject updates from nodes that are not `ready`, forcing clients to heartbeat first. Since this check is done server-side there is no need to coordinate operations client-side: they can continue sending these requests independently and alloc update will succeed after the heartbeat is done. * chagelog: add entry for #15068 * code review * client: skip terminal allocations on reconnect When the client reconnects with the server it synchronizes the state of its allocations by sending data using the `Node.UpdateAlloc` RPC and fetching data using the `Node.GetClientAllocs` RPC. If the data fetch happens before the data write, `unknown` allocations will still be in this state and would trigger the `allocRunner.Reconnect` flow. But when the server `DesiredStatus` for the allocation is `stop` the client should not reconnect the allocation. * apply more code review changes * scheduler: persist changes to reconnected allocs Reconnected allocs have a new AllocState entry that must be persisted by the plan applier. * rpc: read node ID from allocs in UpdateAlloc The AllocUpdateRequest struct is used in three disjoint use cases: 1. Stripped allocs from clients Node.UpdateAlloc RPC using the Allocs, and WriteRequest fields 2. Raft log message using the Allocs, Evals, and WriteRequest fields 3. Plan updates using the AllocsStopped, AllocsUpdated, and Job fields Adding a new field that would only be used in one these cases (1) made things more confusing and error prone. While in theory an AllocUpdateRequest could send allocations from different nodes, in practice this never actually happens since only clients call this method with their own allocations. * scheduler: remove logic to handle exceptional case This condition could only be hit if, somehow, the allocation status was set to "running" while the client was "unknown". This was addressed by enforcing an order in "Node.UpdateStatus" and "Node.UpdateAlloc" RPC calls, so this scenario is not expected to happen. Adding unnecessary code to the scheduler makes it harder to read and reason about it. * more code review * remove another unused test
2022-11-04 20:25:11 +00:00
// Ignore failed allocs that need to be reconnected and that have been
// marked to stop by the server.
if supportsDisconnectedClients &&
Update alloc after reconnect and enforece client heartbeat order (#15068) * scheduler: allow updates after alloc reconnects When an allocation reconnects to a cluster the scheduler needs to run special logic to handle the reconnection, check if a replacement was create and stop one of them. If the allocation kept running while the node was disconnected, it will be reconnected with `ClientStatus: running` and the node will have `Status: ready`. This combination is the same as the normal steady state of allocation, where everything is running as expected. In order to differentiate between the two states (an allocation that is reconnecting and one that is just running) the scheduler needs an extra piece of state. The current implementation uses the presence of a `TaskClientReconnected` task event to detect when the allocation has reconnected and thus must go through the reconnection process. But this event remains even after the allocation is reconnected, causing all future evals to consider the allocation as still reconnecting. This commit changes the reconnect logic to use an `AllocState` to register when the allocation was reconnected. This provides the following benefits: - Only a limited number of task states are kept, and they are used for many other events. It's possible that, upon reconnecting, several actions are triggered that could cause the `TaskClientReconnected` event to be dropped. - Task events are set by clients and so their timestamps are subject to time skew from servers. This prevents using time to determine if an allocation reconnected after a disconnect event. - Disconnect events are already stored as `AllocState` and so storing reconnects there as well makes it the only source of information required. With the new logic, the reconnection logic is only triggered if the last `AllocState` is a disconnect event, meaning that the allocation has not been reconnected yet. After the reconnection is handled, the new `ClientStatus` is store in `AllocState` allowing future evals to skip the reconnection logic. * scheduler: prevent spurious placement on reconnect When a client reconnects it makes two independent RPC calls: - `Node.UpdateStatus` to heartbeat and set its status as `ready`. - `Node.UpdateAlloc` to update the status of its allocations. These two calls can happen in any order, and in case the allocations are updated before a heartbeat it causes the state to be the same as a node being disconnected: the node status will still be `disconnected` while the allocation `ClientStatus` is set to `running`. The current implementation did not handle this order of events properly, and the scheduler would create an unnecessary placement since it considered the allocation was being disconnected. This extra allocation would then be quickly stopped by the heartbeat eval. This commit adds a new code path to handle this order of events. If the node is `disconnected` and the allocation `ClientStatus` is `running` the scheduler will check if the allocation is actually reconnecting using its `AllocState` events. * rpc: only allow alloc updates from `ready` nodes Clients interact with servers using three main RPC methods: - `Node.GetAllocs` reads allocation data from the server and writes it to the client. - `Node.UpdateAlloc` reads allocation from from the client and writes them to the server. - `Node.UpdateStatus` writes the client status to the server and is used as the heartbeat mechanism. These three methods are called periodically by the clients and are done so independently from each other, meaning that there can't be any assumptions in their ordering. This can generate scenarios that are hard to reason about and to code for. For example, when a client misses too many heartbeats it will be considered `down` or `disconnected` and the allocations it was running are set to `lost` or `unknown`. When connectivity is restored the to rest of the cluster, the natural mental model is to think that the client will heartbeat first and then update its allocations status into the servers. But since there's no inherit order in these calls the reverse is just as possible: the client updates the alloc status and then heartbeats. This results in a state where allocs are, for example, `running` while the client is still `disconnected`. This commit adds a new verification to the `Node.UpdateAlloc` method to reject updates from nodes that are not `ready`, forcing clients to heartbeat first. Since this check is done server-side there is no need to coordinate operations client-side: they can continue sending these requests independently and alloc update will succeed after the heartbeat is done. * chagelog: add entry for #15068 * code review * client: skip terminal allocations on reconnect When the client reconnects with the server it synchronizes the state of its allocations by sending data using the `Node.UpdateAlloc` RPC and fetching data using the `Node.GetClientAllocs` RPC. If the data fetch happens before the data write, `unknown` allocations will still be in this state and would trigger the `allocRunner.Reconnect` flow. But when the server `DesiredStatus` for the allocation is `stop` the client should not reconnect the allocation. * apply more code review changes * scheduler: persist changes to reconnected allocs Reconnected allocs have a new AllocState entry that must be persisted by the plan applier. * rpc: read node ID from allocs in UpdateAlloc The AllocUpdateRequest struct is used in three disjoint use cases: 1. Stripped allocs from clients Node.UpdateAlloc RPC using the Allocs, and WriteRequest fields 2. Raft log message using the Allocs, Evals, and WriteRequest fields 3. Plan updates using the AllocsStopped, AllocsUpdated, and Job fields Adding a new field that would only be used in one these cases (1) made things more confusing and error prone. While in theory an AllocUpdateRequest could send allocations from different nodes, in practice this never actually happens since only clients call this method with their own allocations. * scheduler: remove logic to handle exceptional case This condition could only be hit if, somehow, the allocation status was set to "running" while the client was "unknown". This was addressed by enforcing an order in "Node.UpdateStatus" and "Node.UpdateAlloc" RPC calls, so this scenario is not expected to happen. Adding unnecessary code to the scheduler makes it harder to read and reason about it. * more code review * remove another unused test
2022-11-04 20:25:11 +00:00
reconnect &&
alloc.ClientStatus == structs.AllocClientStatusFailed &&
alloc.DesiredStatus == structs.AllocDesiredStatusStop {
ignore[alloc.ID] = alloc
continue
}
if !nodeIsTainted {
// Filter allocs on a node that is now re-connected to be resumed.
Update alloc after reconnect and enforece client heartbeat order (#15068) * scheduler: allow updates after alloc reconnects When an allocation reconnects to a cluster the scheduler needs to run special logic to handle the reconnection, check if a replacement was create and stop one of them. If the allocation kept running while the node was disconnected, it will be reconnected with `ClientStatus: running` and the node will have `Status: ready`. This combination is the same as the normal steady state of allocation, where everything is running as expected. In order to differentiate between the two states (an allocation that is reconnecting and one that is just running) the scheduler needs an extra piece of state. The current implementation uses the presence of a `TaskClientReconnected` task event to detect when the allocation has reconnected and thus must go through the reconnection process. But this event remains even after the allocation is reconnected, causing all future evals to consider the allocation as still reconnecting. This commit changes the reconnect logic to use an `AllocState` to register when the allocation was reconnected. This provides the following benefits: - Only a limited number of task states are kept, and they are used for many other events. It's possible that, upon reconnecting, several actions are triggered that could cause the `TaskClientReconnected` event to be dropped. - Task events are set by clients and so their timestamps are subject to time skew from servers. This prevents using time to determine if an allocation reconnected after a disconnect event. - Disconnect events are already stored as `AllocState` and so storing reconnects there as well makes it the only source of information required. With the new logic, the reconnection logic is only triggered if the last `AllocState` is a disconnect event, meaning that the allocation has not been reconnected yet. After the reconnection is handled, the new `ClientStatus` is store in `AllocState` allowing future evals to skip the reconnection logic. * scheduler: prevent spurious placement on reconnect When a client reconnects it makes two independent RPC calls: - `Node.UpdateStatus` to heartbeat and set its status as `ready`. - `Node.UpdateAlloc` to update the status of its allocations. These two calls can happen in any order, and in case the allocations are updated before a heartbeat it causes the state to be the same as a node being disconnected: the node status will still be `disconnected` while the allocation `ClientStatus` is set to `running`. The current implementation did not handle this order of events properly, and the scheduler would create an unnecessary placement since it considered the allocation was being disconnected. This extra allocation would then be quickly stopped by the heartbeat eval. This commit adds a new code path to handle this order of events. If the node is `disconnected` and the allocation `ClientStatus` is `running` the scheduler will check if the allocation is actually reconnecting using its `AllocState` events. * rpc: only allow alloc updates from `ready` nodes Clients interact with servers using three main RPC methods: - `Node.GetAllocs` reads allocation data from the server and writes it to the client. - `Node.UpdateAlloc` reads allocation from from the client and writes them to the server. - `Node.UpdateStatus` writes the client status to the server and is used as the heartbeat mechanism. These three methods are called periodically by the clients and are done so independently from each other, meaning that there can't be any assumptions in their ordering. This can generate scenarios that are hard to reason about and to code for. For example, when a client misses too many heartbeats it will be considered `down` or `disconnected` and the allocations it was running are set to `lost` or `unknown`. When connectivity is restored the to rest of the cluster, the natural mental model is to think that the client will heartbeat first and then update its allocations status into the servers. But since there's no inherit order in these calls the reverse is just as possible: the client updates the alloc status and then heartbeats. This results in a state where allocs are, for example, `running` while the client is still `disconnected`. This commit adds a new verification to the `Node.UpdateAlloc` method to reject updates from nodes that are not `ready`, forcing clients to heartbeat first. Since this check is done server-side there is no need to coordinate operations client-side: they can continue sending these requests independently and alloc update will succeed after the heartbeat is done. * chagelog: add entry for #15068 * code review * client: skip terminal allocations on reconnect When the client reconnects with the server it synchronizes the state of its allocations by sending data using the `Node.UpdateAlloc` RPC and fetching data using the `Node.GetClientAllocs` RPC. If the data fetch happens before the data write, `unknown` allocations will still be in this state and would trigger the `allocRunner.Reconnect` flow. But when the server `DesiredStatus` for the allocation is `stop` the client should not reconnect the allocation. * apply more code review changes * scheduler: persist changes to reconnected allocs Reconnected allocs have a new AllocState entry that must be persisted by the plan applier. * rpc: read node ID from allocs in UpdateAlloc The AllocUpdateRequest struct is used in three disjoint use cases: 1. Stripped allocs from clients Node.UpdateAlloc RPC using the Allocs, and WriteRequest fields 2. Raft log message using the Allocs, Evals, and WriteRequest fields 3. Plan updates using the AllocsStopped, AllocsUpdated, and Job fields Adding a new field that would only be used in one these cases (1) made things more confusing and error prone. While in theory an AllocUpdateRequest could send allocations from different nodes, in practice this never actually happens since only clients call this method with their own allocations. * scheduler: remove logic to handle exceptional case This condition could only be hit if, somehow, the allocation status was set to "running" while the client was "unknown". This was addressed by enforcing an order in "Node.UpdateStatus" and "Node.UpdateAlloc" RPC calls, so this scenario is not expected to happen. Adding unnecessary code to the scheduler makes it harder to read and reason about it. * more code review * remove another unused test
2022-11-04 20:25:11 +00:00
if reconnect {
if expired {
lost[alloc.ID] = alloc
continue
}
reconnecting[alloc.ID] = alloc
continue
}
// Otherwise, Node is untainted so alloc is untainted
untainted[alloc.ID] = alloc
continue
}
// Allocs on GC'd (nil) or lost nodes are Lost
if taintedNode == nil || taintedNode.TerminalStatus() {
lost[alloc.ID] = alloc
continue
}
// All other allocs are untainted
untainted[alloc.ID] = alloc
}
return
}
// filterByRescheduleable filters the allocation set to return the set of allocations that are either
// untainted or a set of allocations that must be rescheduled now. Allocations that can be rescheduled
// at a future time are also returned so that we can create follow up evaluations for them. Allocs are
// skipped or considered untainted according to logic defined in shouldFilter method.
func (a allocSet) filterByRescheduleable(isBatch, isDisconnecting bool, now time.Time, evalID string, deployment *structs.Deployment) (allocSet, allocSet, []*delayedRescheduleInfo) {
untainted := make(map[string]*structs.Allocation)
rescheduleNow := make(map[string]*structs.Allocation)
rescheduleLater := []*delayedRescheduleInfo{}
for _, alloc := range a {
// Ignore disconnecting allocs that are already unknown. This can happen
// in the case of canaries that are interrupted by a disconnect.
if isDisconnecting && alloc.ClientStatus == structs.AllocClientStatusUnknown {
continue
}
var eligibleNow, eligibleLater bool
var rescheduleTime time.Time
// Ignore failing allocs that have already been rescheduled.
// Only failed or disconnecting allocs should be rescheduled.
// Protects against a bug allowing rescheduling running allocs.
if alloc.NextAllocation != "" && alloc.TerminalStatus() {
continue
2017-05-23 00:14:38 +00:00
}
isUntainted, ignore := shouldFilter(alloc, isBatch)
if isUntainted && !isDisconnecting {
untainted[alloc.ID] = alloc
}
if ignore {
continue
}
eligibleNow, eligibleLater, rescheduleTime = updateByReschedulable(alloc, now, evalID, deployment, isDisconnecting)
if eligibleNow {
rescheduleNow[alloc.ID] = alloc
continue
}
// If the failed alloc is not eligible for rescheduling now we
// add it to the untainted set.
untainted[alloc.ID] = alloc
if eligibleLater {
rescheduleLater = append(rescheduleLater, &delayedRescheduleInfo{alloc.ID, alloc, rescheduleTime})
}
2017-05-23 00:14:38 +00:00
}
return untainted, rescheduleNow, rescheduleLater
}
// shouldFilter returns whether the alloc should be ignored or considered untainted.
//
// Ignored allocs are filtered out.
// Untainted allocs count against the desired total.
// Filtering logic for batch jobs:
// If complete, and ran successfully - untainted
// If desired state is stop - ignore
//
// Filtering logic for service jobs:
// If desired state is stop/evict - ignore
// If client status is complete/lost - ignore
func shouldFilter(alloc *structs.Allocation, isBatch bool) (untainted, ignore bool) {
// Allocs from batch jobs should be filtered when the desired status
// is terminal and the client did not finish or when the client
// status is failed so that they will be replaced. If they are
// complete but not failed, they shouldn't be replaced.
if isBatch {
switch alloc.DesiredStatus {
case structs.AllocDesiredStatusStop:
if alloc.RanSuccessfully() {
return true, false
}
return false, true
case structs.AllocDesiredStatusEvict:
return false, true
}
switch alloc.ClientStatus {
case structs.AllocClientStatusFailed:
return false, false
}
return true, false
}
// Handle service jobs
switch alloc.DesiredStatus {
case structs.AllocDesiredStatusStop, structs.AllocDesiredStatusEvict:
return false, true
}
switch alloc.ClientStatus {
case structs.AllocClientStatusComplete, structs.AllocClientStatusLost:
return false, true
}
return false, false
}
// updateByReschedulable is a helper method that encapsulates logic for whether a failed allocation
// should be rescheduled now, later or left in the untainted set
func updateByReschedulable(alloc *structs.Allocation, now time.Time, evalID string, d *structs.Deployment, isDisconnecting bool) (rescheduleNow, rescheduleLater bool, rescheduleTime time.Time) {
// If the allocation is part of an ongoing active deployment, we only allow it to reschedule
// if it has been marked eligible
if d != nil && alloc.DeploymentID == d.ID && d.Active() && !alloc.DesiredTransition.ShouldReschedule() {
return
}
// Check if the allocation is marked as it should be force rescheduled
if alloc.DesiredTransition.ShouldForceReschedule() {
rescheduleNow = true
}
// Reschedule if the eval ID matches the alloc's followup evalID or if its close to its reschedule time
var eligible bool
switch {
case isDisconnecting:
rescheduleTime, eligible = alloc.NextRescheduleTimeByTime(now)
case alloc.ClientStatus == structs.AllocClientStatusUnknown && alloc.FollowupEvalID == evalID:
lastDisconnectTime := alloc.LastUnknown()
rescheduleTime, eligible = alloc.NextRescheduleTimeByTime(lastDisconnectTime)
default:
rescheduleTime, eligible = alloc.NextRescheduleTime()
}
if eligible && (alloc.FollowupEvalID == evalID || rescheduleTime.Sub(now) <= rescheduleWindowSize) {
rescheduleNow = true
return
}
if eligible && (alloc.FollowupEvalID == "" || isDisconnecting) {
rescheduleLater = true
}
2017-05-23 00:14:38 +00:00
return
}
2018-01-19 21:20:00 +00:00
// filterByTerminal filters out terminal allocs
func filterByTerminal(untainted allocSet) (nonTerminal allocSet) {
nonTerminal = make(map[string]*structs.Allocation)
for id, alloc := range untainted {
if !alloc.TerminalStatus() {
nonTerminal[id] = alloc
}
}
return
}
2017-05-23 00:14:38 +00:00
// filterByDeployment filters allocations into two sets, those that match the
// given deployment ID and those that don't
func (a allocSet) filterByDeployment(id string) (match, nonmatch allocSet) {
match = make(map[string]*structs.Allocation)
nonmatch = make(map[string]*structs.Allocation)
for _, alloc := range a {
if alloc.DeploymentID == id {
match[alloc.ID] = alloc
} else {
nonmatch[alloc.ID] = alloc
}
}
return
}
2017-06-06 21:08:46 +00:00
// delayByStopAfterClientDisconnect returns a delay for any lost allocation that's got a
// stop_after_client_disconnect configured
func (a allocSet) delayByStopAfterClientDisconnect() (later []*delayedRescheduleInfo) {
now := time.Now().UTC()
for _, a := range a {
if !a.ShouldClientStop() {
continue
}
t := a.WaitClientStop()
if t.After(now) {
later = append(later, &delayedRescheduleInfo{
allocID: a.ID,
alloc: a,
rescheduleTime: t,
})
}
}
return later
}
// delayByMaxClientDisconnect returns a delay for any unknown allocation
// that's got a max_client_reconnect configured
func (a allocSet) delayByMaxClientDisconnect(now time.Time) ([]*delayedRescheduleInfo, error) {
var later []*delayedRescheduleInfo
for _, alloc := range a {
timeout := alloc.DisconnectTimeout(now)
if !timeout.After(now) {
return nil, errors.New("unable to computing disconnecting timeouts")
}
later = append(later, &delayedRescheduleInfo{
allocID: alloc.ID,
alloc: alloc,
rescheduleTime: timeout,
})
}
return later, nil
}
// filterOutByClientStatus returns all allocs from the set without the specified client status.
func (a allocSet) filterOutByClientStatus(clientStatus string) allocSet {
allocs := make(allocSet)
for _, alloc := range a {
if alloc.ClientStatus != clientStatus {
allocs[alloc.ID] = alloc
}
}
return allocs
}
// filterByClientStatus returns allocs from the set with the specified client status.
func (a allocSet) filterByClientStatus(clientStatus string) allocSet {
allocs := make(allocSet)
for _, alloc := range a {
if alloc.ClientStatus == clientStatus {
allocs[alloc.ID] = alloc
}
}
return allocs
}
2017-06-06 21:08:46 +00:00
// allocNameIndex is used to select allocation names for placement or removal
// given an existing set of placed allocations.
type allocNameIndex struct {
job, taskGroup string
count int
b structs.Bitmap
// duplicates is used to store duplicate allocation indexes which are
// currently present within the index tracking. The map key is the index,
// and the current count of duplicates. The map is only accessed within a
// single routine and multiple times per job scheduler invocation,
// therefore no lock is used.
duplicates map[uint]int
2017-06-06 21:08:46 +00:00
}
// newAllocNameIndex returns an allocNameIndex for use in selecting names of
// allocations to create or stop. It takes the job and task group name, desired
// count and any existing allocations as input.
func newAllocNameIndex(job, taskGroup string, count int, in allocSet) *allocNameIndex {
bitMap, duplicates := bitmapFrom(in, uint(count))
2017-06-06 21:08:46 +00:00
return &allocNameIndex{
count: count,
b: bitMap,
job: job,
taskGroup: taskGroup,
duplicates: duplicates,
2017-06-06 21:08:46 +00:00
}
}
// bitmapFrom creates a bitmap from the given allocation set and a minimum size
// maybe given. The size of the bitmap is as the larger of the passed minimum
2017-08-15 19:27:05 +00:00
// and the maximum alloc index of the passed input (byte aligned).
func bitmapFrom(input allocSet, minSize uint) (structs.Bitmap, map[uint]int) {
2017-06-06 21:08:46 +00:00
var max uint
for _, a := range input {
if num := a.Index(); num > max {
max = num
}
}
if l := uint(len(input)); minSize < l {
minSize = l
}
2017-06-06 21:08:46 +00:00
if max < minSize {
max = minSize
2017-08-15 19:27:05 +00:00
} else if max%8 == 0 {
// This may be possible if the job was scaled down. We want to make sure
2017-09-26 22:26:33 +00:00
// that the max index is not byte-aligned otherwise we will overflow
// the bitmap.
max++
2017-06-06 21:08:46 +00:00
}
2017-06-06 21:08:46 +00:00
if max == 0 {
max = 8
}
// byteAlign the count
if remainder := max % 8; remainder != 0 {
max = max + 8 - remainder
}
bitmap, err := structs.NewBitmap(max)
if err != nil {
panic(err)
}
// Initialize our duplicates mapping, allowing us to store a non-nil map
// at the cost of 48 bytes.
duplicates := make(map[uint]int)
// Iterate through the allocSet input and hydrate the bitmap. We check that
// the bitmap does not contain the index first, so we can track duplicate
// entries.
2017-06-06 21:08:46 +00:00
for _, a := range input {
allocIndex := a.Index()
if bitmap.Check(allocIndex) {
duplicates[allocIndex]++
} else {
bitmap.Set(allocIndex)
}
2017-06-06 21:08:46 +00:00
}
return bitmap, duplicates
2017-06-06 21:08:46 +00:00
}
// Highest removes and returns the highest n used names. The returned set
2017-06-06 21:08:46 +00:00
// can be less than n if there aren't n names set in the index
func (a *allocNameIndex) Highest(n uint) map[string]struct{} {
h := make(map[string]struct{}, n)
for i := a.b.Size(); i > uint(0) && uint(len(h)) < n; i-- {
// Use this to avoid wrapping around b/c of the unsigned int
idx := i - 1
if a.b.Check(idx) {
a.b.Unset(idx)
h[structs.AllocName(a.job, a.taskGroup, idx)] = struct{}{}
}
}
return h
}
// IsDuplicate checks whether the passed allocation index is duplicated within
// the tracking.
func (a *allocNameIndex) IsDuplicate(idx uint) bool {
val, ok := a.duplicates[idx]
return ok && val > 0
}
2017-06-06 21:08:46 +00:00
// UnsetIndex unsets the index as having its name used
func (a *allocNameIndex) UnsetIndex(idx uint) {
// If this index is a duplicate, remove the duplicate entry. Otherwise, we
// can remove it from the bitmap tracking.
if num, ok := a.duplicates[idx]; ok {
if num--; num == 0 {
delete(a.duplicates, idx)
}
} else {
a.b.Unset(idx)
}
2017-06-06 21:08:46 +00:00
}
// NextCanaries returns the next n names for use as canaries and sets them as
// used. The existing canaries and destructive updates are also passed in.
func (a *allocNameIndex) NextCanaries(n uint, existing, destructive allocSet) []string {
next := make([]string, 0, n)
// Create a name index
existingNames := existing.nameSet()
// First select indexes from the allocations that are undergoing
// destructive updates. This way we avoid duplicate names as they will get
// replaced. As this process already takes into account duplicate checking,
// we can discard the duplicate mapping when generating the bitmap.
dmap, _ := bitmapFrom(destructive, uint(a.count))
remainder := n
2017-06-06 21:08:46 +00:00
for _, idx := range dmap.IndexesInRange(true, uint(0), uint(a.count)-1) {
name := structs.AllocName(a.job, a.taskGroup, uint(idx))
if _, used := existingNames[name]; !used {
next = append(next, name)
a.b.Set(uint(idx))
// If we have enough, return
remainder = n - uint(len(next))
2017-06-06 21:08:46 +00:00
if remainder == 0 {
return next
}
}
}
// Get the set of unset names that can be used
for _, idx := range a.b.IndexesInRange(false, uint(0), uint(a.count)-1) {
name := structs.AllocName(a.job, a.taskGroup, uint(idx))
if _, used := existingNames[name]; !used {
next = append(next, name)
a.b.Set(uint(idx))
// If we have enough, return
remainder = n - uint(len(next))
if remainder == 0 {
return next
}
}
}
// We have exhausted the preferred and free set. Pick starting from n to
2018-04-25 22:03:30 +00:00
// n+remainder, to avoid overlapping where possible. An example is the
// desired count is 3 and we want 5 canaries. The first 3 canaries can use
// index [0, 1, 2] but after that we prefer picking indexes [4, 5] so that
// we do not overlap. Once the canaries are promoted, these would be the
// allocations that would be shut down as well.
for i := uint(a.count); i < uint(a.count)+remainder; i++ {
2017-06-06 21:08:46 +00:00
name := structs.AllocName(a.job, a.taskGroup, i)
next = append(next, name)
2017-06-06 21:08:46 +00:00
}
return next
}
// Next returns the next n names for use as new placements and sets them as
// used.
func (a *allocNameIndex) Next(n uint) []string {
next := make([]string, 0, n)
// Get the set of unset names that can be used
remainder := n
for _, idx := range a.b.IndexesInRange(false, uint(0), uint(a.count)-1) {
next = append(next, structs.AllocName(a.job, a.taskGroup, uint(idx)))
a.b.Set(uint(idx))
// If we have enough, return
remainder = n - uint(len(next))
if remainder == 0 {
return next
}
}
// We have exhausted the free set, now just pick overlapping indexes
var i uint
for i = 0; i < remainder; i++ {
next = append(next, structs.AllocName(a.job, a.taskGroup, i))
a.b.Set(i)
}
return next
}