open-consul/agent/consul/config_replication.go

236 lines
7.1 KiB
Go
Raw Normal View History

package consul
import (
"context"
"fmt"
"sort"
"time"
"github.com/armon/go-metrics"
"github.com/hashicorp/go-hclog"
"github.com/hashicorp/consul/agent/structs"
)
func configSort(configs []structs.ConfigEntry) {
server: config entry replication now correctly uses namespaces in comparisons (#9024) Previously config entries sharing a kind & name but in different namespaces could occasionally cause "stuck states" in replication because the namespace fields were ignored during the differential comparison phase. Example: Two config entries written to the primary: kind=A,name=web,namespace=bar kind=A,name=web,namespace=foo Under the covers these both get saved to memdb, so they are sorted by all 3 components (kind,name,namespace) during natural iteration. This means that before the replication code does it's own incomplete sort, the underlying data IS sorted by namespace ascending (bar comes before foo). After one pass of replication the primary and secondary datacenters have the same set of config entries present. If "kind=A,name=web,namespace=bar" were to be deleted, then things get weird. Before replication the two sides look like: primary: [ kind=A,name=web,namespace=foo ] secondary: [ kind=A,name=web,namespace=bar kind=A,name=web,namespace=foo ] The differential comparison phase walks these two lists in sorted order and first compares "kind=A,name=web,namespace=foo" vs "kind=A,name=web,namespace=bar" and falsely determines they are the SAME and are thus cause an update of "kind=A,name=web,namespace=foo". Then it compares "<nothing>" with "kind=A,name=web,namespace=foo" and falsely determines that the latter should be DELETED. During reconciliation the deletes are processed before updates, and so for a brief moment in the secondary "kind=A,name=web,namespace=foo" is erroneously deleted and then immediately restored. Unfortunately after this replication phase the final state is identical to the initial state, so when it loops around again (rate limited) it repeats the same set of operations indefinitely.
2020-10-23 18:41:54 +00:00
sort.SliceStable(configs, func(i, j int) bool {
return cmpConfigLess(configs[i], configs[j])
})
}
func diffConfigEntries(local []structs.ConfigEntry, remote []structs.ConfigEntry, lastRemoteIndex uint64) ([]structs.ConfigEntry, []structs.ConfigEntry) {
configSort(local)
configSort(remote)
server: config entry replication now correctly uses namespaces in comparisons (#9024) Previously config entries sharing a kind & name but in different namespaces could occasionally cause "stuck states" in replication because the namespace fields were ignored during the differential comparison phase. Example: Two config entries written to the primary: kind=A,name=web,namespace=bar kind=A,name=web,namespace=foo Under the covers these both get saved to memdb, so they are sorted by all 3 components (kind,name,namespace) during natural iteration. This means that before the replication code does it's own incomplete sort, the underlying data IS sorted by namespace ascending (bar comes before foo). After one pass of replication the primary and secondary datacenters have the same set of config entries present. If "kind=A,name=web,namespace=bar" were to be deleted, then things get weird. Before replication the two sides look like: primary: [ kind=A,name=web,namespace=foo ] secondary: [ kind=A,name=web,namespace=bar kind=A,name=web,namespace=foo ] The differential comparison phase walks these two lists in sorted order and first compares "kind=A,name=web,namespace=foo" vs "kind=A,name=web,namespace=bar" and falsely determines they are the SAME and are thus cause an update of "kind=A,name=web,namespace=foo". Then it compares "<nothing>" with "kind=A,name=web,namespace=foo" and falsely determines that the latter should be DELETED. During reconciliation the deletes are processed before updates, and so for a brief moment in the secondary "kind=A,name=web,namespace=foo" is erroneously deleted and then immediately restored. Unfortunately after this replication phase the final state is identical to the initial state, so when it loops around again (rate limited) it repeats the same set of operations indefinitely.
2020-10-23 18:41:54 +00:00
var (
deletions []structs.ConfigEntry
updates []structs.ConfigEntry
localIdx int
remoteIdx int
)
for localIdx, remoteIdx = 0, 0; localIdx < len(local) && remoteIdx < len(remote); {
server: config entry replication now correctly uses namespaces in comparisons (#9024) Previously config entries sharing a kind & name but in different namespaces could occasionally cause "stuck states" in replication because the namespace fields were ignored during the differential comparison phase. Example: Two config entries written to the primary: kind=A,name=web,namespace=bar kind=A,name=web,namespace=foo Under the covers these both get saved to memdb, so they are sorted by all 3 components (kind,name,namespace) during natural iteration. This means that before the replication code does it's own incomplete sort, the underlying data IS sorted by namespace ascending (bar comes before foo). After one pass of replication the primary and secondary datacenters have the same set of config entries present. If "kind=A,name=web,namespace=bar" were to be deleted, then things get weird. Before replication the two sides look like: primary: [ kind=A,name=web,namespace=foo ] secondary: [ kind=A,name=web,namespace=bar kind=A,name=web,namespace=foo ] The differential comparison phase walks these two lists in sorted order and first compares "kind=A,name=web,namespace=foo" vs "kind=A,name=web,namespace=bar" and falsely determines they are the SAME and are thus cause an update of "kind=A,name=web,namespace=foo". Then it compares "<nothing>" with "kind=A,name=web,namespace=foo" and falsely determines that the latter should be DELETED. During reconciliation the deletes are processed before updates, and so for a brief moment in the secondary "kind=A,name=web,namespace=foo" is erroneously deleted and then immediately restored. Unfortunately after this replication phase the final state is identical to the initial state, so when it loops around again (rate limited) it repeats the same set of operations indefinitely.
2020-10-23 18:41:54 +00:00
if configSameID(local[localIdx], remote[remoteIdx]) {
// config is in both the local and remote state - need to check raft indices
if remote[remoteIdx].GetRaftIndex().ModifyIndex > lastRemoteIndex {
updates = append(updates, remote[remoteIdx])
}
// increment both indices when equal
localIdx += 1
remoteIdx += 1
} else if cmpConfigLess(local[localIdx], remote[remoteIdx]) {
// config no longer in remoted state - needs deleting
deletions = append(deletions, local[localIdx])
// increment just the local index
localIdx += 1
} else {
// local state doesn't have this config - needs updating
updates = append(updates, remote[remoteIdx])
// increment just the remote index
remoteIdx += 1
}
}
for ; localIdx < len(local); localIdx += 1 {
deletions = append(deletions, local[localIdx])
}
for ; remoteIdx < len(remote); remoteIdx += 1 {
updates = append(updates, remote[remoteIdx])
}
return deletions, updates
}
server: config entry replication now correctly uses namespaces in comparisons (#9024) Previously config entries sharing a kind & name but in different namespaces could occasionally cause "stuck states" in replication because the namespace fields were ignored during the differential comparison phase. Example: Two config entries written to the primary: kind=A,name=web,namespace=bar kind=A,name=web,namespace=foo Under the covers these both get saved to memdb, so they are sorted by all 3 components (kind,name,namespace) during natural iteration. This means that before the replication code does it's own incomplete sort, the underlying data IS sorted by namespace ascending (bar comes before foo). After one pass of replication the primary and secondary datacenters have the same set of config entries present. If "kind=A,name=web,namespace=bar" were to be deleted, then things get weird. Before replication the two sides look like: primary: [ kind=A,name=web,namespace=foo ] secondary: [ kind=A,name=web,namespace=bar kind=A,name=web,namespace=foo ] The differential comparison phase walks these two lists in sorted order and first compares "kind=A,name=web,namespace=foo" vs "kind=A,name=web,namespace=bar" and falsely determines they are the SAME and are thus cause an update of "kind=A,name=web,namespace=foo". Then it compares "<nothing>" with "kind=A,name=web,namespace=foo" and falsely determines that the latter should be DELETED. During reconciliation the deletes are processed before updates, and so for a brief moment in the secondary "kind=A,name=web,namespace=foo" is erroneously deleted and then immediately restored. Unfortunately after this replication phase the final state is identical to the initial state, so when it loops around again (rate limited) it repeats the same set of operations indefinitely.
2020-10-23 18:41:54 +00:00
func cmpConfigLess(first structs.ConfigEntry, second structs.ConfigEntry) bool {
if first.GetKind() < second.GetKind() {
return true
}
if first.GetKind() > second.GetKind() {
return false
}
if first.GetEnterpriseMeta().LessThan(second.GetEnterpriseMeta()) {
return true
}
if second.GetEnterpriseMeta().LessThan(first.GetEnterpriseMeta()) {
return false
}
return first.GetName() < second.GetName()
}
func configSameID(e1, e2 structs.ConfigEntry) bool {
return e1.GetKind() == e2.GetKind() &&
e1.GetEnterpriseMeta().IsSame(e2.GetEnterpriseMeta()) &&
e1.GetName() == e2.GetName()
}
func (s *Server) reconcileLocalConfig(ctx context.Context, configs []structs.ConfigEntry, op structs.ConfigEntryOp) (bool, error) {
ticker := time.NewTicker(time.Second / time.Duration(s.config.ConfigReplicationApplyLimit))
defer ticker.Stop()
for i, entry := range configs {
req := structs.ConfigEntryRequest{
Op: op,
Datacenter: s.config.Datacenter,
Entry: entry,
}
_, err := s.raftApply(structs.ConfigEntryRequestType, &req)
if err != nil {
return false, fmt.Errorf("Failed to apply config %s: %v", op, err)
}
2020-08-21 15:42:37 +00:00
if i < len(configs)-1 {
select {
case <-ctx.Done():
return true, nil
case <-ticker.C:
// do nothing - ready for the next batch
}
}
}
return false, nil
}
func (s *Server) fetchConfigEntries(lastRemoteIndex uint64) (*structs.IndexedGenericConfigEntries, error) {
defer metrics.MeasureSince([]string{"leader", "replication", "config-entries", "fetch"}, time.Now())
connect: intentions are now managed as a new config entry kind "service-intentions" (#8834) - Upgrade the ConfigEntry.ListAll RPC to be kind-aware so that older copies of consul will not see new config entries it doesn't understand replicate down. - Add shim conversion code so that the old API/CLI method of interacting with intentions will continue to work so long as none of these are edited via config entry endpoints. Almost all of the read-only APIs will continue to function indefinitely. - Add new APIs that operate on individual intentions without IDs so that the UI doesn't need to implement CAS operations. - Add a new serf feature flag indicating support for intentions-as-config-entries. - The old line-item intentions way of interacting with the state store will transparently flip between the legacy memdb table and the config entry representations so that readers will never see a hiccup during migration where the results are incomplete. It uses a piece of system metadata to control the flip. - The primary datacenter will begin migrating intentions into config entries on startup once all servers in the datacenter are on a version of Consul with the intentions-as-config-entries feature flag. When it is complete the old state store representations will be cleared. We also record a piece of system metadata indicating this has occurred. We use this metadata to skip ALL of this code the next time the leader starts up. - The secondary datacenters continue to run the old intentions replicator until all servers in the secondary DC and primary DC support intentions-as-config-entries (via serf flag). Once this condition it met the old intentions replicator ceases. - The secondary datacenters replicate the new config entries as they are migrated in the primary. When they detect that the primary has zeroed it's old state store table it waits until all config entries up to that point are replicated and then zeroes its own copy of the old state store table. We also record a piece of system metadata indicating this has occurred. We use this metadata to skip ALL of this code the next time the leader starts up.
2020-10-06 18:24:05 +00:00
req := structs.ConfigEntryListAllRequest{
Datacenter: s.config.PrimaryDatacenter,
connect: intentions are now managed as a new config entry kind "service-intentions" (#8834) - Upgrade the ConfigEntry.ListAll RPC to be kind-aware so that older copies of consul will not see new config entries it doesn't understand replicate down. - Add shim conversion code so that the old API/CLI method of interacting with intentions will continue to work so long as none of these are edited via config entry endpoints. Almost all of the read-only APIs will continue to function indefinitely. - Add new APIs that operate on individual intentions without IDs so that the UI doesn't need to implement CAS operations. - Add a new serf feature flag indicating support for intentions-as-config-entries. - The old line-item intentions way of interacting with the state store will transparently flip between the legacy memdb table and the config entry representations so that readers will never see a hiccup during migration where the results are incomplete. It uses a piece of system metadata to control the flip. - The primary datacenter will begin migrating intentions into config entries on startup once all servers in the datacenter are on a version of Consul with the intentions-as-config-entries feature flag. When it is complete the old state store representations will be cleared. We also record a piece of system metadata indicating this has occurred. We use this metadata to skip ALL of this code the next time the leader starts up. - The secondary datacenters continue to run the old intentions replicator until all servers in the secondary DC and primary DC support intentions-as-config-entries (via serf flag). Once this condition it met the old intentions replicator ceases. - The secondary datacenters replicate the new config entries as they are migrated in the primary. When they detect that the primary has zeroed it's old state store table it waits until all config entries up to that point are replicated and then zeroes its own copy of the old state store table. We also record a piece of system metadata indicating this has occurred. We use this metadata to skip ALL of this code the next time the leader starts up.
2020-10-06 18:24:05 +00:00
Kinds: structs.AllConfigEntryKinds,
QueryOptions: structs.QueryOptions{
AllowStale: true,
MinQueryIndex: lastRemoteIndex,
Token: s.tokens.ReplicationToken(),
},
EnterpriseMeta: *s.replicationEnterpriseMeta(),
}
var response structs.IndexedGenericConfigEntries
if err := s.RPC("ConfigEntry.ListAll", &req, &response); err != nil {
return nil, err
}
return &response, nil
}
func (s *Server) replicateConfig(ctx context.Context, lastRemoteIndex uint64, logger hclog.Logger) (uint64, bool, error) {
remote, err := s.fetchConfigEntries(lastRemoteIndex)
if err != nil {
return 0, false, fmt.Errorf("failed to retrieve remote config entries: %v", err)
}
logger.Debug("finished fetching config entries", "amount", len(remote.Entries))
// Need to check if we should be stopping. This will be common as the fetching process is a blocking
// RPC which could have been hanging around for a long time and during that time leadership could
// have been lost.
select {
case <-ctx.Done():
return 0, true, nil
default:
// do nothing
}
// Measure everything after the remote query, which can block for long
// periods of time. This metric is a good measure of how expensive the
// replication process is.
defer metrics.MeasureSince([]string{"leader", "replication", "config", "apply"}, time.Now())
_, local, err := s.fsm.State().ConfigEntries(nil, s.replicationEnterpriseMeta())
if err != nil {
return 0, false, fmt.Errorf("failed to retrieve local config entries: %v", err)
}
// If the remote index ever goes backwards, it's a good indication that
// the remote side was rebuilt and we should do a full sync since we
// can't make any assumptions about what's going on.
//
// Resetting lastRemoteIndex to 0 will work because we never consider local
// raft indices. Instead we compare the raft modify index in the response object
// with the lastRemoteIndex (only when we already have a config entry of the same kind/name)
// to determine if an update is needed. Resetting lastRemoteIndex to 0 then has the affect
// of making us think all the local state is out of date and any matching entries should
// still be updated.
//
// The lastRemoteIndex is not used when the entry exists either only in the local state or
// only in the remote state. In those situations we need to either delete it or create it.
if remote.QueryMeta.Index < lastRemoteIndex {
logger.Warn("Config Entry replication remote index moved backwards, forcing a full Config Entry sync",
"from", lastRemoteIndex,
"to", remote.QueryMeta.Index,
)
lastRemoteIndex = 0
}
logger.Debug("Config Entry replication",
"local", len(local),
"remote", len(remote.Entries),
)
// Calculate the changes required to bring the state into sync and then
// apply them.
deletions, updates := diffConfigEntries(local, remote.Entries, lastRemoteIndex)
logger.Debug("Config Entry replication",
"deletions", len(deletions),
"updates", len(updates),
)
if len(deletions) > 0 {
logger.Debug("Deleting local config entries",
"deletions", len(deletions),
)
exit, err := s.reconcileLocalConfig(ctx, deletions, structs.ConfigEntryDelete)
if exit {
return 0, true, nil
}
if err != nil {
return 0, false, fmt.Errorf("failed to delete local config entries: %v", err)
}
logger.Debug("Config Entry replication - finished deletions")
}
if len(updates) > 0 {
logger.Debug("Updating local config entries",
"updates", len(updates),
)
exit, err := s.reconcileLocalConfig(ctx, updates, structs.ConfigEntryUpsert)
if exit {
return 0, true, nil
}
if err != nil {
return 0, false, fmt.Errorf("failed to update local config entries: %v", err)
}
logger.Debug("Config Entry replication - finished updates")
}
// Return the index we got back from the remote side, since we've synced
// up with the remote state as of that index.
return remote.QueryMeta.Index, false, nil
}