open-nomad/testutil/wait.go

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

361 lines
9.7 KiB
Go
Raw Normal View History

2015-06-04 11:02:39 +00:00
package testutil
2015-06-04 11:26:16 +00:00
import (
"fmt"
"os"
"runtime"
"testing"
2015-06-04 11:26:16 +00:00
"time"
core: enforce strict steps for clients reconnect (#15808) When a Nomad client that is running an allocation with `max_client_disconnect` set misses a heartbeat the Nomad server will update its status to `disconnected`. Upon reconnecting, the client will make three main RPC calls: - `Node.UpdateStatus` is used to set the client status to `ready`. - `Node.UpdateAlloc` is used to update the client-side information about allocations, such as their `ClientStatus`, task states etc. - `Node.Register` is used to upsert the entire node information, including its status. These calls are made concurrently and are also running in parallel with the scheduler. Depending on the order they run the scheduler may end up with incomplete data when reconciling allocations. For example, a client disconnects and its replacement allocation cannot be placed anywhere else, so there's a pending eval waiting for resources. When this client comes back the order of events may be: 1. Client calls `Node.UpdateStatus` and is now `ready`. 2. Scheduler reconciles allocations and places the replacement alloc to the client. The client is now assigned two allocations: the original alloc that is still `unknown` and the replacement that is `pending`. 3. Client calls `Node.UpdateAlloc` and updates the original alloc to `running`. 4. Scheduler notices too many allocs and stops the replacement. This creates unnecessary placements or, in a different order of events, may leave the job without any allocations running until the whole state is updated and reconciled. To avoid problems like this clients must update _all_ of its relevant information before they can be considered `ready` and available for scheduling. To achieve this goal the RPC endpoints mentioned above have been modified to enforce strict steps for nodes reconnecting: - `Node.Register` does not set the client status anymore. - `Node.UpdateStatus` sets the reconnecting client to the `initializing` status until it successfully calls `Node.UpdateAlloc`. These changes are done server-side to avoid the need of additional coordination between clients and servers. Clients are kept oblivious of these changes and will keep making these calls as they normally would. The verification of whether allocations have been updates is done by storing and comparing the Raft index of the last time the client missed a heartbeat and the last time it updated its allocations.
2023-01-25 20:53:59 +00:00
"github.com/google/go-cmp/cmp"
"github.com/hashicorp/nomad/nomad/structs"
"github.com/kr/pretty"
core: enforce strict steps for clients reconnect (#15808) When a Nomad client that is running an allocation with `max_client_disconnect` set misses a heartbeat the Nomad server will update its status to `disconnected`. Upon reconnecting, the client will make three main RPC calls: - `Node.UpdateStatus` is used to set the client status to `ready`. - `Node.UpdateAlloc` is used to update the client-side information about allocations, such as their `ClientStatus`, task states etc. - `Node.Register` is used to upsert the entire node information, including its status. These calls are made concurrently and are also running in parallel with the scheduler. Depending on the order they run the scheduler may end up with incomplete data when reconciling allocations. For example, a client disconnects and its replacement allocation cannot be placed anywhere else, so there's a pending eval waiting for resources. When this client comes back the order of events may be: 1. Client calls `Node.UpdateStatus` and is now `ready`. 2. Scheduler reconciles allocations and places the replacement alloc to the client. The client is now assigned two allocations: the original alloc that is still `unknown` and the replacement that is `pending`. 3. Client calls `Node.UpdateAlloc` and updates the original alloc to `running`. 4. Scheduler notices too many allocs and stops the replacement. This creates unnecessary placements or, in a different order of events, may leave the job without any allocations running until the whole state is updated and reconciled. To avoid problems like this clients must update _all_ of its relevant information before they can be considered `ready` and available for scheduling. To achieve this goal the RPC endpoints mentioned above have been modified to enforce strict steps for nodes reconnecting: - `Node.Register` does not set the client status anymore. - `Node.UpdateStatus` sets the reconnecting client to the `initializing` status until it successfully calls `Node.UpdateAlloc`. These changes are done server-side to avoid the need of additional coordination between clients and servers. Clients are kept oblivious of these changes and will keep making these calls as they normally would. The verification of whether allocations have been updates is done by storing and comparing the Raft index of the last time the client missed a heartbeat and the last time it updated its allocations.
2023-01-25 20:53:59 +00:00
"github.com/shoenig/test/must"
2015-06-04 11:26:16 +00:00
)
2015-06-04 11:02:39 +00:00
type testFn func() (bool, error)
type errorFn func(error)
func Wait(t *testing.T, test testFn) {
t.Helper()
retries := 500 * TestMultiplier()
warn := int64(float64(retries) * 0.75)
for tries := retries; tries > 0; {
time.Sleep(10 * time.Millisecond)
tries--
success, err := test()
if success {
return
}
switch tries {
case 0:
if err == nil {
t.Fatalf("timeout waiting for test function to succeed (you should probably return a helpful error instead of nil!)")
} else {
t.Fatalf("timeout: %v", err)
}
case warn:
pc, _, _, _ := runtime.Caller(1)
f := runtime.FuncForPC(pc)
t.Logf("%d/%d retries reached for %s (err=%v)", warn, retries, f.Name(), err)
}
}
}
2015-06-04 11:02:39 +00:00
func WaitForResult(test testFn, error errorFn) {
WaitForResultRetries(500*TestMultiplier(), test, error)
}
2015-06-04 11:02:39 +00:00
2016-01-21 20:29:13 +00:00
func WaitForResultRetries(retries int64, test testFn, error errorFn) {
2015-06-04 11:02:39 +00:00
for retries > 0 {
time.Sleep(10 * time.Millisecond)
retries--
success, err := test()
if success {
return
}
if retries == 0 {
error(err)
}
}
}
2015-06-04 11:26:16 +00:00
// WaitForResultUntil waits the duration for the test to pass.
// Otherwise error is called after the deadline expires.
func WaitForResultUntil(until time.Duration, test testFn, errorFunc errorFn) {
var success bool
var err error
deadline := time.Now().Add(until)
for time.Now().Before(deadline) {
success, err = test()
if success {
return
}
// Sleep some arbitrary fraction of the deadline
time.Sleep(until / 30)
}
errorFunc(err)
}
2017-04-07 00:05:09 +00:00
// AssertUntil asserts the test function passes throughout the given duration.
// Otherwise error is called on failure.
func AssertUntil(until time.Duration, test testFn, error errorFn) {
deadline := time.Now().Add(until)
for time.Now().Before(deadline) {
success, err := test()
if !success {
error(err)
return
}
// Sleep some arbitrary fraction of the deadline
time.Sleep(until / 30)
}
}
// TestMultiplier returns a multiplier for retries and waits given environment
// the tests are being run under.
2016-01-21 20:29:13 +00:00
func TestMultiplier() int64 {
if IsCI() {
return 4
}
return 1
}
2018-04-03 23:51:57 +00:00
// Timeout takes the desired timeout and increases it if running in Travis
func Timeout(original time.Duration) time.Duration {
return original * time.Duration(TestMultiplier())
}
func IsCI() bool {
_, ok := os.LookupEnv("CI")
return ok
}
2016-01-21 21:28:48 +00:00
func IsTravis() bool {
_, ok := os.LookupEnv("TRAVIS")
return ok
}
func IsAppVeyor() bool {
_, ok := os.LookupEnv("APPVEYOR")
2016-01-21 21:28:48 +00:00
return ok
}
2015-06-04 11:26:16 +00:00
type rpcFn func(string, interface{}, interface{}) error
// WaitForLeader blocks until a leader is elected.
func WaitForLeader(t testing.TB, rpc rpcFn) {
t.Helper()
2015-06-04 11:26:16 +00:00
WaitForResult(func() (bool, error) {
args := &structs.GenericRequest{}
2015-06-04 11:26:16 +00:00
var leader string
err := rpc("Status.Leader", args, &leader)
return leader != "", err
}, func(err error) {
t.Fatalf("failed to find leader: %v", err)
})
}
// WaitForClient blocks until the client can be found
func WaitForClient(t testing.TB, rpc rpcFn, nodeID string, region string) {
core: enforce strict steps for clients reconnect (#15808) When a Nomad client that is running an allocation with `max_client_disconnect` set misses a heartbeat the Nomad server will update its status to `disconnected`. Upon reconnecting, the client will make three main RPC calls: - `Node.UpdateStatus` is used to set the client status to `ready`. - `Node.UpdateAlloc` is used to update the client-side information about allocations, such as their `ClientStatus`, task states etc. - `Node.Register` is used to upsert the entire node information, including its status. These calls are made concurrently and are also running in parallel with the scheduler. Depending on the order they run the scheduler may end up with incomplete data when reconciling allocations. For example, a client disconnects and its replacement allocation cannot be placed anywhere else, so there's a pending eval waiting for resources. When this client comes back the order of events may be: 1. Client calls `Node.UpdateStatus` and is now `ready`. 2. Scheduler reconciles allocations and places the replacement alloc to the client. The client is now assigned two allocations: the original alloc that is still `unknown` and the replacement that is `pending`. 3. Client calls `Node.UpdateAlloc` and updates the original alloc to `running`. 4. Scheduler notices too many allocs and stops the replacement. This creates unnecessary placements or, in a different order of events, may leave the job without any allocations running until the whole state is updated and reconciled. To avoid problems like this clients must update _all_ of its relevant information before they can be considered `ready` and available for scheduling. To achieve this goal the RPC endpoints mentioned above have been modified to enforce strict steps for nodes reconnecting: - `Node.Register` does not set the client status anymore. - `Node.UpdateStatus` sets the reconnecting client to the `initializing` status until it successfully calls `Node.UpdateAlloc`. These changes are done server-side to avoid the need of additional coordination between clients and servers. Clients are kept oblivious of these changes and will keep making these calls as they normally would. The verification of whether allocations have been updates is done by storing and comparing the Raft index of the last time the client missed a heartbeat and the last time it updated its allocations.
2023-01-25 20:53:59 +00:00
t.Helper()
WaitForClientStatus(t, rpc, nodeID, region, structs.NodeStatusReady)
}
core: enforce strict steps for clients reconnect (#15808) When a Nomad client that is running an allocation with `max_client_disconnect` set misses a heartbeat the Nomad server will update its status to `disconnected`. Upon reconnecting, the client will make three main RPC calls: - `Node.UpdateStatus` is used to set the client status to `ready`. - `Node.UpdateAlloc` is used to update the client-side information about allocations, such as their `ClientStatus`, task states etc. - `Node.Register` is used to upsert the entire node information, including its status. These calls are made concurrently and are also running in parallel with the scheduler. Depending on the order they run the scheduler may end up with incomplete data when reconciling allocations. For example, a client disconnects and its replacement allocation cannot be placed anywhere else, so there's a pending eval waiting for resources. When this client comes back the order of events may be: 1. Client calls `Node.UpdateStatus` and is now `ready`. 2. Scheduler reconciles allocations and places the replacement alloc to the client. The client is now assigned two allocations: the original alloc that is still `unknown` and the replacement that is `pending`. 3. Client calls `Node.UpdateAlloc` and updates the original alloc to `running`. 4. Scheduler notices too many allocs and stops the replacement. This creates unnecessary placements or, in a different order of events, may leave the job without any allocations running until the whole state is updated and reconciled. To avoid problems like this clients must update _all_ of its relevant information before they can be considered `ready` and available for scheduling. To achieve this goal the RPC endpoints mentioned above have been modified to enforce strict steps for nodes reconnecting: - `Node.Register` does not set the client status anymore. - `Node.UpdateStatus` sets the reconnecting client to the `initializing` status until it successfully calls `Node.UpdateAlloc`. These changes are done server-side to avoid the need of additional coordination between clients and servers. Clients are kept oblivious of these changes and will keep making these calls as they normally would. The verification of whether allocations have been updates is done by storing and comparing the Raft index of the last time the client missed a heartbeat and the last time it updated its allocations.
2023-01-25 20:53:59 +00:00
// WaitForClientStatus blocks until the client is in the expected status.
func WaitForClientStatus(t testing.TB, rpc rpcFn, nodeID string, region string, status string) {
t.Helper()
if region == "" {
region = "global"
}
WaitForResult(func() (bool, error) {
req := structs.NodeSpecificRequest{
NodeID: nodeID,
QueryOptions: structs.QueryOptions{Region: region},
}
var out structs.SingleNodeResponse
err := rpc("Node.GetNode", &req, &out)
if err != nil {
return false, err
}
if out.Node == nil {
return false, fmt.Errorf("node not found")
}
core: enforce strict steps for clients reconnect (#15808) When a Nomad client that is running an allocation with `max_client_disconnect` set misses a heartbeat the Nomad server will update its status to `disconnected`. Upon reconnecting, the client will make three main RPC calls: - `Node.UpdateStatus` is used to set the client status to `ready`. - `Node.UpdateAlloc` is used to update the client-side information about allocations, such as their `ClientStatus`, task states etc. - `Node.Register` is used to upsert the entire node information, including its status. These calls are made concurrently and are also running in parallel with the scheduler. Depending on the order they run the scheduler may end up with incomplete data when reconciling allocations. For example, a client disconnects and its replacement allocation cannot be placed anywhere else, so there's a pending eval waiting for resources. When this client comes back the order of events may be: 1. Client calls `Node.UpdateStatus` and is now `ready`. 2. Scheduler reconciles allocations and places the replacement alloc to the client. The client is now assigned two allocations: the original alloc that is still `unknown` and the replacement that is `pending`. 3. Client calls `Node.UpdateAlloc` and updates the original alloc to `running`. 4. Scheduler notices too many allocs and stops the replacement. This creates unnecessary placements or, in a different order of events, may leave the job without any allocations running until the whole state is updated and reconciled. To avoid problems like this clients must update _all_ of its relevant information before they can be considered `ready` and available for scheduling. To achieve this goal the RPC endpoints mentioned above have been modified to enforce strict steps for nodes reconnecting: - `Node.Register` does not set the client status anymore. - `Node.UpdateStatus` sets the reconnecting client to the `initializing` status until it successfully calls `Node.UpdateAlloc`. These changes are done server-side to avoid the need of additional coordination between clients and servers. Clients are kept oblivious of these changes and will keep making these calls as they normally would. The verification of whether allocations have been updates is done by storing and comparing the Raft index of the last time the client missed a heartbeat and the last time it updated its allocations.
2023-01-25 20:53:59 +00:00
if out.Node.Status != status {
return false, fmt.Errorf("node is %s, not %s", out.Node.Status, status)
}
return true, nil
}, func(err error) {
core: enforce strict steps for clients reconnect (#15808) When a Nomad client that is running an allocation with `max_client_disconnect` set misses a heartbeat the Nomad server will update its status to `disconnected`. Upon reconnecting, the client will make three main RPC calls: - `Node.UpdateStatus` is used to set the client status to `ready`. - `Node.UpdateAlloc` is used to update the client-side information about allocations, such as their `ClientStatus`, task states etc. - `Node.Register` is used to upsert the entire node information, including its status. These calls are made concurrently and are also running in parallel with the scheduler. Depending on the order they run the scheduler may end up with incomplete data when reconciling allocations. For example, a client disconnects and its replacement allocation cannot be placed anywhere else, so there's a pending eval waiting for resources. When this client comes back the order of events may be: 1. Client calls `Node.UpdateStatus` and is now `ready`. 2. Scheduler reconciles allocations and places the replacement alloc to the client. The client is now assigned two allocations: the original alloc that is still `unknown` and the replacement that is `pending`. 3. Client calls `Node.UpdateAlloc` and updates the original alloc to `running`. 4. Scheduler notices too many allocs and stops the replacement. This creates unnecessary placements or, in a different order of events, may leave the job without any allocations running until the whole state is updated and reconciled. To avoid problems like this clients must update _all_ of its relevant information before they can be considered `ready` and available for scheduling. To achieve this goal the RPC endpoints mentioned above have been modified to enforce strict steps for nodes reconnecting: - `Node.Register` does not set the client status anymore. - `Node.UpdateStatus` sets the reconnecting client to the `initializing` status until it successfully calls `Node.UpdateAlloc`. These changes are done server-side to avoid the need of additional coordination between clients and servers. Clients are kept oblivious of these changes and will keep making these calls as they normally would. The verification of whether allocations have been updates is done by storing and comparing the Raft index of the last time the client missed a heartbeat and the last time it updated its allocations.
2023-01-25 20:53:59 +00:00
t.Fatalf("failed to wait for node staus: %v", err)
})
core: enforce strict steps for clients reconnect (#15808) When a Nomad client that is running an allocation with `max_client_disconnect` set misses a heartbeat the Nomad server will update its status to `disconnected`. Upon reconnecting, the client will make three main RPC calls: - `Node.UpdateStatus` is used to set the client status to `ready`. - `Node.UpdateAlloc` is used to update the client-side information about allocations, such as their `ClientStatus`, task states etc. - `Node.Register` is used to upsert the entire node information, including its status. These calls are made concurrently and are also running in parallel with the scheduler. Depending on the order they run the scheduler may end up with incomplete data when reconciling allocations. For example, a client disconnects and its replacement allocation cannot be placed anywhere else, so there's a pending eval waiting for resources. When this client comes back the order of events may be: 1. Client calls `Node.UpdateStatus` and is now `ready`. 2. Scheduler reconciles allocations and places the replacement alloc to the client. The client is now assigned two allocations: the original alloc that is still `unknown` and the replacement that is `pending`. 3. Client calls `Node.UpdateAlloc` and updates the original alloc to `running`. 4. Scheduler notices too many allocs and stops the replacement. This creates unnecessary placements or, in a different order of events, may leave the job without any allocations running until the whole state is updated and reconciled. To avoid problems like this clients must update _all_ of its relevant information before they can be considered `ready` and available for scheduling. To achieve this goal the RPC endpoints mentioned above have been modified to enforce strict steps for nodes reconnecting: - `Node.Register` does not set the client status anymore. - `Node.UpdateStatus` sets the reconnecting client to the `initializing` status until it successfully calls `Node.UpdateAlloc`. These changes are done server-side to avoid the need of additional coordination between clients and servers. Clients are kept oblivious of these changes and will keep making these calls as they normally would. The verification of whether allocations have been updates is done by storing and comparing the Raft index of the last time the client missed a heartbeat and the last time it updated its allocations.
2023-01-25 20:53:59 +00:00
t.Logf("[TEST] Client for test %s %s, id: %s, region: %s", t.Name(), status, nodeID, region)
}
// WaitForVotingMembers blocks until autopilot promotes all server peers
// to be voting members.
//
// Useful for tests that change cluster topology (e.g. kill a node)
// that should wait until cluster is stable.
func WaitForVotingMembers(t testing.TB, rpc rpcFn, nPeers int) {
WaitForResult(func() (bool, error) {
args := &structs.GenericRequest{}
args.AllowStale = true
args.Region = "global"
args.Namespace = structs.DefaultNamespace
resp := structs.RaftConfigurationResponse{}
err := rpc("Operator.RaftGetConfiguration", args, &resp)
if err != nil {
return false, fmt.Errorf("failed to query raft: %v", err)
}
if len(resp.Servers) != nPeers {
return false, fmt.Errorf("expected %d peers found %d", nPeers, len(resp.Servers))
}
for _, s := range resp.Servers {
if !s.Voter {
return false, fmt.Errorf("found nonvoting server: %v", s)
}
}
return true, nil
}, func(err error) {
t.Fatalf("failed to wait until voting members: %v", err)
})
}
// RegisterJobWithToken registers a job and uses the job's Region and Namespace.
func RegisterJobWithToken(t testing.TB, rpc rpcFn, job *structs.Job, token string) {
WaitForResult(func() (bool, error) {
args := &structs.JobRegisterRequest{}
args.Job = job
args.WriteRequest.Region = job.Region
args.AuthToken = token
args.Namespace = job.Namespace
var jobResp structs.JobRegisterResponse
err := rpc("Job.Register", args, &jobResp)
return err == nil, fmt.Errorf("Job.Register error: %v", err)
}, func(err error) {
t.Fatalf("error registering job: %v", err)
})
t.Logf("Job %q registered", job.ID)
}
func RegisterJob(t testing.TB, rpc rpcFn, job *structs.Job) {
RegisterJobWithToken(t, rpc, job, "")
}
func WaitForRunningWithToken(t testing.TB, rpc rpcFn, job *structs.Job, token string) []*structs.AllocListStub {
RegisterJobWithToken(t, rpc, job, token)
var resp structs.JobAllocationsResponse
// This can be quite slow if the job has expensive setup such as
// downloading large artifacts or creating a chroot.
WaitForResultRetries(2000*TestMultiplier(), func() (bool, error) {
args := &structs.JobSpecificRequest{}
args.JobID = job.ID
args.QueryOptions.Region = job.Region
args.AuthToken = token
args.Namespace = job.Namespace
err := rpc("Job.Allocations", args, &resp)
if err != nil {
return false, fmt.Errorf("Job.Allocations error: %v", err)
}
if len(resp.Allocations) == 0 {
evals := structs.JobEvaluationsResponse{}
must.NoError(t, rpc("Job.Evaluations", args, &evals), must.Sprintf("error looking up evals"))
return false, fmt.Errorf("0 allocations; evals: %s", pretty.Sprint(evals.Evaluations))
}
for _, alloc := range resp.Allocations {
if alloc.ClientStatus == structs.AllocClientStatusPending {
return false, fmt.Errorf("alloc not running: id=%v tg=%v status=%v",
alloc.ID, alloc.TaskGroup, alloc.ClientStatus)
}
}
return true, nil
}, func(err error) {
must.NoError(t, err)
})
return resp.Allocations
}
// WaitForRunning runs a job and blocks until all allocs are out of pending.
func WaitForRunning(t testing.TB, rpc rpcFn, job *structs.Job) []*structs.AllocListStub {
return WaitForRunningWithToken(t, rpc, job, "")
}
core: enforce strict steps for clients reconnect (#15808) When a Nomad client that is running an allocation with `max_client_disconnect` set misses a heartbeat the Nomad server will update its status to `disconnected`. Upon reconnecting, the client will make three main RPC calls: - `Node.UpdateStatus` is used to set the client status to `ready`. - `Node.UpdateAlloc` is used to update the client-side information about allocations, such as their `ClientStatus`, task states etc. - `Node.Register` is used to upsert the entire node information, including its status. These calls are made concurrently and are also running in parallel with the scheduler. Depending on the order they run the scheduler may end up with incomplete data when reconciling allocations. For example, a client disconnects and its replacement allocation cannot be placed anywhere else, so there's a pending eval waiting for resources. When this client comes back the order of events may be: 1. Client calls `Node.UpdateStatus` and is now `ready`. 2. Scheduler reconciles allocations and places the replacement alloc to the client. The client is now assigned two allocations: the original alloc that is still `unknown` and the replacement that is `pending`. 3. Client calls `Node.UpdateAlloc` and updates the original alloc to `running`. 4. Scheduler notices too many allocs and stops the replacement. This creates unnecessary placements or, in a different order of events, may leave the job without any allocations running until the whole state is updated and reconciled. To avoid problems like this clients must update _all_ of its relevant information before they can be considered `ready` and available for scheduling. To achieve this goal the RPC endpoints mentioned above have been modified to enforce strict steps for nodes reconnecting: - `Node.Register` does not set the client status anymore. - `Node.UpdateStatus` sets the reconnecting client to the `initializing` status until it successfully calls `Node.UpdateAlloc`. These changes are done server-side to avoid the need of additional coordination between clients and servers. Clients are kept oblivious of these changes and will keep making these calls as they normally would. The verification of whether allocations have been updates is done by storing and comparing the Raft index of the last time the client missed a heartbeat and the last time it updated its allocations.
2023-01-25 20:53:59 +00:00
// WaitforJobAllocStatus blocks until the ClientStatus of allocations for a job
// match the expected map of <ClientStatus>: <count>.
func WaitForJobAllocStatus(t testing.TB, rpc rpcFn, job *structs.Job, allocStatus map[string]int) {
t.Helper()
WaitForJobAllocStatusWithToken(t, rpc, job, allocStatus, "")
}
// WaitForJobAllocStatusWithToken behaves the same way as WaitForJobAllocStatus
// but is used for clusters with ACL enabled.
func WaitForJobAllocStatusWithToken(t testing.TB, rpc rpcFn, job *structs.Job, allocStatus map[string]int, token string) []*structs.AllocListStub {
core: enforce strict steps for clients reconnect (#15808) When a Nomad client that is running an allocation with `max_client_disconnect` set misses a heartbeat the Nomad server will update its status to `disconnected`. Upon reconnecting, the client will make three main RPC calls: - `Node.UpdateStatus` is used to set the client status to `ready`. - `Node.UpdateAlloc` is used to update the client-side information about allocations, such as their `ClientStatus`, task states etc. - `Node.Register` is used to upsert the entire node information, including its status. These calls are made concurrently and are also running in parallel with the scheduler. Depending on the order they run the scheduler may end up with incomplete data when reconciling allocations. For example, a client disconnects and its replacement allocation cannot be placed anywhere else, so there's a pending eval waiting for resources. When this client comes back the order of events may be: 1. Client calls `Node.UpdateStatus` and is now `ready`. 2. Scheduler reconciles allocations and places the replacement alloc to the client. The client is now assigned two allocations: the original alloc that is still `unknown` and the replacement that is `pending`. 3. Client calls `Node.UpdateAlloc` and updates the original alloc to `running`. 4. Scheduler notices too many allocs and stops the replacement. This creates unnecessary placements or, in a different order of events, may leave the job without any allocations running until the whole state is updated and reconciled. To avoid problems like this clients must update _all_ of its relevant information before they can be considered `ready` and available for scheduling. To achieve this goal the RPC endpoints mentioned above have been modified to enforce strict steps for nodes reconnecting: - `Node.Register` does not set the client status anymore. - `Node.UpdateStatus` sets the reconnecting client to the `initializing` status until it successfully calls `Node.UpdateAlloc`. These changes are done server-side to avoid the need of additional coordination between clients and servers. Clients are kept oblivious of these changes and will keep making these calls as they normally would. The verification of whether allocations have been updates is done by storing and comparing the Raft index of the last time the client missed a heartbeat and the last time it updated its allocations.
2023-01-25 20:53:59 +00:00
t.Helper()
var allocs []*structs.AllocListStub
core: enforce strict steps for clients reconnect (#15808) When a Nomad client that is running an allocation with `max_client_disconnect` set misses a heartbeat the Nomad server will update its status to `disconnected`. Upon reconnecting, the client will make three main RPC calls: - `Node.UpdateStatus` is used to set the client status to `ready`. - `Node.UpdateAlloc` is used to update the client-side information about allocations, such as their `ClientStatus`, task states etc. - `Node.Register` is used to upsert the entire node information, including its status. These calls are made concurrently and are also running in parallel with the scheduler. Depending on the order they run the scheduler may end up with incomplete data when reconciling allocations. For example, a client disconnects and its replacement allocation cannot be placed anywhere else, so there's a pending eval waiting for resources. When this client comes back the order of events may be: 1. Client calls `Node.UpdateStatus` and is now `ready`. 2. Scheduler reconciles allocations and places the replacement alloc to the client. The client is now assigned two allocations: the original alloc that is still `unknown` and the replacement that is `pending`. 3. Client calls `Node.UpdateAlloc` and updates the original alloc to `running`. 4. Scheduler notices too many allocs and stops the replacement. This creates unnecessary placements or, in a different order of events, may leave the job without any allocations running until the whole state is updated and reconciled. To avoid problems like this clients must update _all_ of its relevant information before they can be considered `ready` and available for scheduling. To achieve this goal the RPC endpoints mentioned above have been modified to enforce strict steps for nodes reconnecting: - `Node.Register` does not set the client status anymore. - `Node.UpdateStatus` sets the reconnecting client to the `initializing` status until it successfully calls `Node.UpdateAlloc`. These changes are done server-side to avoid the need of additional coordination between clients and servers. Clients are kept oblivious of these changes and will keep making these calls as they normally would. The verification of whether allocations have been updates is done by storing and comparing the Raft index of the last time the client missed a heartbeat and the last time it updated its allocations.
2023-01-25 20:53:59 +00:00
WaitForResultRetries(2000*TestMultiplier(), func() (bool, error) {
args := &structs.JobSpecificRequest{
JobID: job.ID,
QueryOptions: structs.QueryOptions{
AuthToken: token,
Namespace: job.Namespace,
Region: job.Region,
},
}
var resp structs.JobAllocationsResponse
err := rpc("Job.Allocations", args, &resp)
if err != nil {
return false, fmt.Errorf("Job.Allocations error: %v", err)
}
if len(resp.Allocations) == 0 {
evals := structs.JobEvaluationsResponse{}
must.NoError(t, rpc("Job.Evaluations", args, &evals), must.Sprintf("error looking up evals"))
core: enforce strict steps for clients reconnect (#15808) When a Nomad client that is running an allocation with `max_client_disconnect` set misses a heartbeat the Nomad server will update its status to `disconnected`. Upon reconnecting, the client will make three main RPC calls: - `Node.UpdateStatus` is used to set the client status to `ready`. - `Node.UpdateAlloc` is used to update the client-side information about allocations, such as their `ClientStatus`, task states etc. - `Node.Register` is used to upsert the entire node information, including its status. These calls are made concurrently and are also running in parallel with the scheduler. Depending on the order they run the scheduler may end up with incomplete data when reconciling allocations. For example, a client disconnects and its replacement allocation cannot be placed anywhere else, so there's a pending eval waiting for resources. When this client comes back the order of events may be: 1. Client calls `Node.UpdateStatus` and is now `ready`. 2. Scheduler reconciles allocations and places the replacement alloc to the client. The client is now assigned two allocations: the original alloc that is still `unknown` and the replacement that is `pending`. 3. Client calls `Node.UpdateAlloc` and updates the original alloc to `running`. 4. Scheduler notices too many allocs and stops the replacement. This creates unnecessary placements or, in a different order of events, may leave the job without any allocations running until the whole state is updated and reconciled. To avoid problems like this clients must update _all_ of its relevant information before they can be considered `ready` and available for scheduling. To achieve this goal the RPC endpoints mentioned above have been modified to enforce strict steps for nodes reconnecting: - `Node.Register` does not set the client status anymore. - `Node.UpdateStatus` sets the reconnecting client to the `initializing` status until it successfully calls `Node.UpdateAlloc`. These changes are done server-side to avoid the need of additional coordination between clients and servers. Clients are kept oblivious of these changes and will keep making these calls as they normally would. The verification of whether allocations have been updates is done by storing and comparing the Raft index of the last time the client missed a heartbeat and the last time it updated its allocations.
2023-01-25 20:53:59 +00:00
return false, fmt.Errorf("0 allocations; evals: %s", pretty.Sprint(evals.Evaluations))
}
allocs = resp.Allocations
core: enforce strict steps for clients reconnect (#15808) When a Nomad client that is running an allocation with `max_client_disconnect` set misses a heartbeat the Nomad server will update its status to `disconnected`. Upon reconnecting, the client will make three main RPC calls: - `Node.UpdateStatus` is used to set the client status to `ready`. - `Node.UpdateAlloc` is used to update the client-side information about allocations, such as their `ClientStatus`, task states etc. - `Node.Register` is used to upsert the entire node information, including its status. These calls are made concurrently and are also running in parallel with the scheduler. Depending on the order they run the scheduler may end up with incomplete data when reconciling allocations. For example, a client disconnects and its replacement allocation cannot be placed anywhere else, so there's a pending eval waiting for resources. When this client comes back the order of events may be: 1. Client calls `Node.UpdateStatus` and is now `ready`. 2. Scheduler reconciles allocations and places the replacement alloc to the client. The client is now assigned two allocations: the original alloc that is still `unknown` and the replacement that is `pending`. 3. Client calls `Node.UpdateAlloc` and updates the original alloc to `running`. 4. Scheduler notices too many allocs and stops the replacement. This creates unnecessary placements or, in a different order of events, may leave the job without any allocations running until the whole state is updated and reconciled. To avoid problems like this clients must update _all_ of its relevant information before they can be considered `ready` and available for scheduling. To achieve this goal the RPC endpoints mentioned above have been modified to enforce strict steps for nodes reconnecting: - `Node.Register` does not set the client status anymore. - `Node.UpdateStatus` sets the reconnecting client to the `initializing` status until it successfully calls `Node.UpdateAlloc`. These changes are done server-side to avoid the need of additional coordination between clients and servers. Clients are kept oblivious of these changes and will keep making these calls as they normally would. The verification of whether allocations have been updates is done by storing and comparing the Raft index of the last time the client missed a heartbeat and the last time it updated its allocations.
2023-01-25 20:53:59 +00:00
got := map[string]int{}
for _, alloc := range resp.Allocations {
got[alloc.ClientStatus]++
}
if diff := cmp.Diff(allocStatus, got); diff != "" {
return false, fmt.Errorf("alloc status mismatch (-want +got):\n%s", diff)
}
return true, nil
}, func(err error) {
must.NoError(t, err)
})
return allocs
core: enforce strict steps for clients reconnect (#15808) When a Nomad client that is running an allocation with `max_client_disconnect` set misses a heartbeat the Nomad server will update its status to `disconnected`. Upon reconnecting, the client will make three main RPC calls: - `Node.UpdateStatus` is used to set the client status to `ready`. - `Node.UpdateAlloc` is used to update the client-side information about allocations, such as their `ClientStatus`, task states etc. - `Node.Register` is used to upsert the entire node information, including its status. These calls are made concurrently and are also running in parallel with the scheduler. Depending on the order they run the scheduler may end up with incomplete data when reconciling allocations. For example, a client disconnects and its replacement allocation cannot be placed anywhere else, so there's a pending eval waiting for resources. When this client comes back the order of events may be: 1. Client calls `Node.UpdateStatus` and is now `ready`. 2. Scheduler reconciles allocations and places the replacement alloc to the client. The client is now assigned two allocations: the original alloc that is still `unknown` and the replacement that is `pending`. 3. Client calls `Node.UpdateAlloc` and updates the original alloc to `running`. 4. Scheduler notices too many allocs and stops the replacement. This creates unnecessary placements or, in a different order of events, may leave the job without any allocations running until the whole state is updated and reconciled. To avoid problems like this clients must update _all_ of its relevant information before they can be considered `ready` and available for scheduling. To achieve this goal the RPC endpoints mentioned above have been modified to enforce strict steps for nodes reconnecting: - `Node.Register` does not set the client status anymore. - `Node.UpdateStatus` sets the reconnecting client to the `initializing` status until it successfully calls `Node.UpdateAlloc`. These changes are done server-side to avoid the need of additional coordination between clients and servers. Clients are kept oblivious of these changes and will keep making these calls as they normally would. The verification of whether allocations have been updates is done by storing and comparing the Raft index of the last time the client missed a heartbeat and the last time it updated its allocations.
2023-01-25 20:53:59 +00:00
}
// WaitForFiles blocks until all the files in the slice are present
func WaitForFiles(t testing.TB, files []string) {
WaitForResult(func() (bool, error) {
return FilesExist(files)
}, func(err error) {
t.Fatalf("missing expected files: %v", err)
})
}
// WaitForFilesUntil blocks until duration or all the files in the slice are present
func WaitForFilesUntil(t testing.TB, files []string, until time.Duration) {
WaitForResultUntil(until, func() (bool, error) {
return FilesExist(files)
}, func(err error) {
t.Fatalf("missing expected files: %v", err)
})
}
// FilesExist verifies all files in the slice are present
func FilesExist(files []string) (bool, error) {
for _, f := range files {
if _, err := os.Stat(f); os.IsNotExist(err) {
return false, fmt.Errorf("expected file not found: %v", f)
}
}
return true, nil
}