When trace logging is enabled we log the computed steps for a policy. This
commit makes sure that the steps that are logged are in the same order they will
be run when the policy executes. This makes it much easier to reason about the
policy if the move-to-step API is ever required in the future.
It is possible that the Unfollow API may fail to release shard history
retention leases when unfollowing, so this needs to be handled by the
ILM Unfollow action. There's nothing much that can be done automatically
about it from the follower side, so this change makes the ILM unfollow
action simply ignore those failures.
This test should no longer pass when the functionality it is intended to
test is broken, as it now indexes a number of documents and verifies
that the index is staying on the same step until after indexing and
replication of those documents is finished. This prevents the test from
passing if the leader index progresses in its lifecycle during that time.
The SchedulerEngine is used in several places in our code and not all
of these usages properly stopped the SchedulerEngine, which could lead
to test failures due to leaked threads from the SchedulerEngine. This
change adds stopping to these usages in order to avoid the thread leaks
that cause CI failures and noise.
Closes#38875
This commit attempts to remove the retention leases on the leader shards
when unfollowing an index. This is best effort, since the leader might
not be available.
This commit is the first step in integrating shard history retention
leases with CCR. In this commit we integrate shard history retention
leases with recovery from remote. Before we start transferring files, we
take out a retention lease on the primary. Then during the file copy
phase, we repeatedly renew the retention lease. Finally, when recovery
from remote is complete, we disable the background renewing of the
retention lease.
We have had various reports of problems caused by the maxRetryTimeout
setting in the low-level REST client. Such setting was initially added
in the attempts to not have requests go through retries if the request
already took longer than the provided timeout.
The implementation was problematic though as such timeout would also
expire in the first request attempt (see #31834), would leave the
request executing after expiration causing memory leaks (see #33342),
and would not take into account the http client internal queuing (see #25951).
Given all these issues, it seems that this custom timeout mechanism
gives little benefits while causing a lot of harm. We should rather rely
on connect and socket timeout exposed by the underlying http client
and accept that a request can overall take longer than the configured
timeout, which is the case even with a single retry anyways.
This commit removes the `maxRetryTimeout` setting and all of its usages.
`CreateIndexRequest#source(Map<String, Object>, ... )`, which is used when
deserializing index creation requests, accidentally accepts mappings that are
nested twice under the type key (as described in the bug report #38266).
This in turn causes us to be too lenient in parsing typeless mappings. In
particular, we accept the following index creation request, even though it
should not contain the type key `_doc`:
```
PUT index?include_type_name=false
{
"mappings": {
"_doc": {
"properties": { ... }
}
}
}
```
There is a similar issue for both 'put templates' and 'put mappings' requests
as well.
This PR makes the minimal changes to detect and reject these typed mappings in
requests. It does not address #38266 generally, or attempt a larger refactor
around types in these server-side requests, as I think this should be done at a
later time.
This test should not pass until CCR finishes integrating shard history
retention leases. It currently sometimes passes (which is a bug in the
test), but cannot pass reliably until the linked issue is resolved.
Adds a Step to the Shrink and Delete actions which prevents those
actions from running on a leader index - all follower indices must first
unfollow the leader index before these actions can run. This prevents
the loss of history before follower indices are ready, which might
otherwise result in the loss of data.
There was a bug where creating a new policy would start
the ILM service, even if it was stopped. This change ensures
that there is no change to the existing operation mode
Previously, ShrinkAction would fail if
it was executed on an index that had
the same number of shards as the target
shrunken number.
This PR introduced a new BranchingStep that
is used inside of ShrinkAction to branch which
step to move to next, depending on the
shard values. So no shrink will occur if the
shard count is unchanged.
This commit modifies the put follow index action to use a
CcrRepository when creating a follower index. It routes
the logic through the snapshot/restore process. A
wait_for_active_shards parameter can be used to configure
how long to wait before returning the response.
We inject an Unfollow action before Shrink because the Shrink action
cannot be safely used on a following index, as it may not be fully
caught up with the leader index before the "original" following index is
deleted and replaced with a non-following Shrunken index. The Unfollow
action will verify that 1) the index is marked as "complete", and 2) all
operations up to this point have been replicated from the leader to the
follower before explicitly disconnecting the follower from the leader.
Injecting an Unfollow action before the Rollover action is done mainly
as a convenience: This allow users to use the same lifecycle policy on
both the leader and follower cluster without having to explictly modify
the policy to unfollow the index, while doing what we expect users to
want in most cases.
* Use ILM for Watcher history deletion
This commit adds an index lifecycle policy for the `.watch-history-*` indices.
This policy is automatically used for all new watch history indices.
This does not yet remove the automatic cleanup that the monitoring plugin does
for the .watch-history indices, and it does not touch the
`xpack.watcher.history.cleaner_service.enabled` setting.
Relates to #32041
Some steps, such as steps that delete, close, or freeze an index, may fail due to a currently running snapshot of the index. In those cases, rather than move to the ERROR step, we should retry the step when the snapshot has completed.
This change adds an abstract step (`AsyncRetryDuringSnapshotActionStep`) that certain steps (like the ones I mentioned above) can extend that will automatically handle a situation where a snapshot is taking place. When a `SnapshotInProgressException` is received by the listener wrapper, a `ClusterStateObserver` listener is registered to wait until the snapshot has completed, re-running the ILM action when no snapshot is occurring.
This also adds integration tests for these scenarios (thanks to @talevy in #37552).
Resolves#37541
The integ tests currently use the raw zip project name as the
distribution type. This commit simplifies this specification to be
"default" or "oss". Whether zip or tar is used should be an internal
implementation detail of the integ test setup, which can (in the future)
be platform specific.
This change adds the unfollow action for CCR follower indices.
This is needed for the shrink action in case an index is a follower index.
This will give the follower index the opportunity to fully catch up with
the leader index, pause index following and unfollow the leader index.
After this the shrink action can safely perform the ilm shrink.
The unfollow action needs to be added to the hot phase and acts as
barrier for going to the next phase (warm or delete phases), so that
follower indices are being unfollowed properly before indices are expected
to go in read-only mode. This allows the force merge action to execute
its steps safely.
The unfollow action has three steps:
* `wait-for-indexing-complete` step: waits for the index in question
to get the `index.lifecycle.indexing_complete` setting be set to `true`
* `wait-for-follow-shard-tasks` step: waits for all the shard follow tasks
for the index being handled to report that the leader shard global checkpoint
is equal to the follower shard global checkpoint.
* `pause-follower-index` step: Pauses index following, necessary to unfollow
* `close-follower-index` step: Closes the index, necessary to unfollow
* `unfollow-follower-index` step: Actually unfollows the index using
the CCR Unfollow API
* `open-follower-index` step: Reopens the index now that it is a normal index
* `wait-for-yellow` step: Waits for primary shards to be allocated after
reopening the index to ensure the index is ready for the next step
In the case of the last two steps, if the index in being handled is
a regular index then the steps acts as a no-op.
Relates to #34648
Co-authored-by: Martijn van Groningen <martijn.v.groningen@gmail.com>
Co-authored-by: Gordon Brown <gordon.brown@elastic.co>
This commit adds a set_priority action to the hot, warm, and cold
phases for an ILM policy. This action sets the `index.priority`
on the managed index to allow different priorities between the
hot, warm, and cold recoveries.
This commit also includes the HLRC and documentation changes.
closes#36905
* Tests: Add ElasticsearchAssertions.awaitLatch method
Some tests are using assertTrue(latch.await(...)) in their code. This
leads to an assertion error without any error message. This adds a
method which has a nicer error message and can be used in tests.
* fix forbidden apis
* fix spaces
the testFullPolicy and testMoveToRolloverStep tests
are very important tests, but they sometimes timeout
beyond the default 10sec wait for shrink to occur.
This commit increases one of the assertBusys to
20 seconds
Leaving `index.lifecycle.indexing_complete` in place when removing the
lifecycle policy from an index can cause confusion, as if a new policy
is associated with the policy, rollover will be silently skipped.
Removing that setting when removing the policy from an index makes
associating a new policy with the index more involved, but allows ILM to
fail loudly, rather than silently skipping operations which the user may
assume are being performed.
* Adjust order of checks in WaitForRolloverReadyStep
This allows ILM to error out properly for indices that have a valid
alias, but are not the write index, while still handling
`indexing_complete` on old-style aliases and rollover (that is, those
which only point to a single index at a time with no explicit write
index)
* WaitForRolloverReadyStepTests#mutateInstance sometimes did not mutate the instance
correctly
* 40_explain_lifecycle#"Test new phase still has phase_time" is not really a necessary
integration test. In addition to this, it is flaky due to the asynchronous nature of
ILM metadata population
* add read_ilm cluster privilege
Although managing ILM policies is best done using the
"manage" cluster privilege, it is useful to have read-only
views.
* adds `read_ilm` cluster privilege for viewing policies and status
* adds Explain API to the `view_index_metadata` index privilege
* add manage_ilm privileges
The error message used when attempting to delete a lifecycle policy that
is in use previously only included one index which was using the policy.
It now includes all indices using that policy.
Adds a setting that indicates that an index is done indexing, set by ILM
when the Rollover action completes. This indicates that the Rollover
action should be skipped in any future invocations, as long as the index
is no longer the write index for its alias.
This enables 1) an index with a policy that involves the Rollover action
to have the policy removed and switched to another one without use of
the move-to-step API, and 2) integrations with Beats and CCR.
RolloverStep previously had a name of "attempt_rollover", which was
inconsistent with all other step names due it its use of an underscore
instead of a dash.
RolloverAction will now periodically check the rollover conditions using
the Rollover API with the dry_run option as an AsyncWaitStep, then run
the rollover itself by calling the Rollover API with no conditions,
which will always roll over, as an AsyncActionStep. This will resolve
race condition issues in policies using RolloverAction.
Before, moving to a failed step would only change the step info
to be that of the failed step. This means two things.
1. Async Steps would never be triggered to execute
2. If there are inherent problems with the action definition that can
be fixed with a policy update, these changes were not being reflected
by the new execution info.
Changes now
1. Async steps are executed after the move to the failed step in cluster state
2. the lifecycle execution info's phase definition is updated from the current
latest policy definition, even though the index isn't moving to a new phase.
Closes#35397.
With this change, `Version` no longer carries information about the qualifier,
we still need a way to show the "display version" that does have both
qualifier and snapshot. This is now stored by the build and red from `META-INF`.
The remove-ilm-from-index API was using the DELETE http method
to signify that something is being removed. Although, metadata
about ILM for the index is being deleted, no entity/resource
is being deleted during this operation. POST is more in line with
what this API is actually doing, it is modifying the metadata for
an index. As part of this change, `remove` is also appended to the path
to be more explicit about its actions.
If the Rollover step would fail due to the next index in sequence
already existing, just skip to the next step instead of going to the
Error step.
This prevents spurious `ResourceAlreadyExistsException`s created by
simultaneous RolloverStep executions from causing ILM to error out
unnecessarily.
This commit removes the Joda time usage from ILM and the HLRC components of ILM.
It also fixes an issue where using the `?human=true` flag could have caused the
parser not to work. These millisecond fields now follow the standard we use
elsewhere in the code, with additional fields added iff the `human` flag is
specified.
This is a breaking change for ILM, but since ILM has not yet been released, no
compatibility shim is needed.
Previously, testRunStateChangePolicyWithNextStep asserted that the
ClusterState before and after running the steps were equal. The test
only passed due to a race condition: The latch would be triggered by the
step execution, but the cluster state update thread would continue
running before committing the change to the cluster state. This allowed
the test to read the old cluster state and pass the equality check about
99.99% of the time.
The test now waits for the new cluster state to be committed before
checking that it is _not_ equal to the old cluster state.
This commit does a few things
- moves ILM-specifc rest yaml tests into plugin/ilm/qa, and creates special
:plugin:ilm:qa:rest module to test them
- removes the with-security tests of the yaml tests since they are covered in
the rest tests now
- moves ChangePolicyforIndexIT into the qa/multi-node project since that test is
not currently running in main ilm since integTest is disabled
This PR renames the CRUD APIS for ILM
GET _ilm/<policy>, _ilm -> _ilm/policy/<policy>, _ilm/policy
PUT _ilm/<policy> -> _ilm/policy/<policy>
DELETE _ilm/<policy> -> _ilm/policy/<policy>
closes#34929.
Previously, if ClusterStateActionSteps or ClusterStateWaitSteps threw an
exception executing, the exception would only be caught and logged by
the generic ClusterStateUpdateTask machinery and the index would become
stuck on that step.
Now, exceptions thrown in these steps will be caught and the index will
be moved to the Error step.
The Move To Step API now checks to see if the target step is an
AsyncActionStep, and if so, runs it.
Previously, AsyncActionSteps would only be run when they are entered by
executing the previous step, so if an AsyncActionStep was entered via
the Move To Step API, ILM would never touch that index again.
ILM would return a resource-not-found exception when requesting policies
while the IndexLifecycleMetaData is not initialized. The behavior here
should not be as extreme since it is not the user's fault.
This commit changes the behavior so that it succeeds and returns no policies
when no policy names are explicitely specified, otherwise keep the same behavior
of throwing an exception
With the introduction of _ilm/stop and _ilm/start APIs, the
use cases where one would only target a select group
of indices to start/stop has been reduced. Since there is no
strong use-case for skipping specific indices, it is best to
remove this functionality and only adding if later desired, with the
hopes of keeping things more simple.
through randomization, there is a chance that the mutateInstance
for PolicyStatsTests does not actually mutate the original object.
This PR aims to fix this
* Adds usage data for ILM
* Adds tests for IndexLifecycleFeatureSetUsage and friends
* Adds tests for IndexLifecycleFeatureSet
* Fixes merge errors
* Add number of indices managed to usage stats
Also adds more tests
* Addresses Review comments
* Removes Set Policy API in favour of setting index.lifecycle.name
directly
* Reinstates matcher that will still be used
* Cleans up code after rebase
* Adds test to check changing policy with ndex settings works
* Fixes TimeseriesLifecycleActionsIT after API removal
* Fixes docs tests
* Fixes case on close where lifecycle service was never created
Since there's no transition into the "new" phase it wasn't set until the "hot"
phase, so now we initialize it when initializing the policy context.
Resolves#34277
Having integration tests separated from the unit tests in the qa
directory works much more smoothly with our testing infrastructure,
matches what other plugins do, and tests in a more "real" deployment
scenario by having all plugins installed.
This change adds the command RemoveIndexLifecyclePolicy to the HLRC. This uses the
new TimeRequest as a base class for RemoveIndexLifecyclePolicyRequest on the client side.
This fixes an issue where an incorrect expected next step is used when checking
to execute `AsyncActionStep`s after a cluster state step.
It fixes this scenario:
- `ExecuteStepsUpdateTask` executes a `ClusterStateWaitStep` or
`ClusterStateActionStep` successfully
- The next step is also a `ClusterStateWaitStep`, so it loops
- The `ClusterStateWaitStep` has a next stepkey (which gets set to the
`nextStepKey` in the code)
- The `ClusterStateWaitStep` fails the condition, meaning that it will have to
wait longer
- The `nextStepKey` is now incorrect though, because we did not advance the
index's step, and it's not `null` (which is another safe value if there is no
step after the `ClusterStateWaitStep`)
This fixes the problem by resetting the nextStepKey to null if the condition is
not met, since we are not going to advance the step metadata in this
case (thereby skipping the `maybeRunAsyncAction` invocation).
This commit also tightens up and enhances much of the ILM logging. A lot of
logging was missing the index name (making it hard to debug in the presence of
multiple indices) and a lot was using the wrong logging level (DEBUG is now
actually readable without being a wall of text).
Resolves#34297
- this adds an integration test that runs through a policy
with all the actions defined.
- adds a test specific to a policy having just a rollover action
- bumps the node count to 4
This commit changes the way that step execution flows. Rather than have any step
run when the cluster state changes or the periodic scheduler fires, this now
runs the different types of steps at different times.
`AsyncWaitStep` is run at a periodic manner, ie, every 10 minutes by default
`ClusterStateActionStep` and `ClusterStateWaitStep` are run every time the
cluster state changes.
`AsyncActionStep` is now run only after the cluster state has been transitioned
into a new step. This prevents these non-idempotent steps from running at the
same time. It addition to being run when transitioned into, this is also run
when a node is newly elected master (only if set as the current step) so that
master failover does not fail to run the step.
This also changes the `RolloverStep` from an `AsyncActionStep` to an
`AsyncWaitStep` so that it can run periodically.
Relates to #29823
ILM now only forces indices to become read only in the case of Shrink
and Force Merge actions, as these are most useful in cases where the
index is no longer being written to.
Using index settings for ILM state is fragile and exposes too much
information that doesn't need to be exposed. Using custom index metadata
is more resilient and allows more controlled access to internal
information.
As part of these changes, moves away from using defaults for ILM-related
values, in favor of using null values to clearly indicate that the value is not
present.
This moves away from caching a list of steps for a current phase, instead
rebuilding the necessary step from the phase JSON stored in the index's
metadata.
Relates to #29823
There was a listener that re-runs the policy with the new state when the cluster
state is processed by the `MoveToNextStepUpdateTask`. This removes this listener
as we will execute the policy through the `IndexLifecyleService` cluster state
listener.
It is useful to keep track of which version of a policy is currently
being executed by a specific index. For management purposes, it would
also be useful to know at which time the latest version was inserted
so that an audit trail is left for reconciling changes happening in ILM.
This removes `PhaseAfterStep` in favor of a new `PhaseCompleteStep`. This step
in only a marker that the `LifecyclePolicyRunner` needs to halt until the time
indicated for entering the next phase.
This also fixes a bug where phase times were encapsulated into the policy
instead of dynamically adjusting to policy changes.
Supersedes #33140, which it replaces
Relates to #29823
Since policies can be updated independent of execution plans for the current
phase being executed, it would be nice to know what the phase that is executing
looks like in JSON. This PR does just that, while also using that index setting
to recontruct the phase steps to execute (for consistency)
This brings the name in line with everywhere else and means that name
seen on the feature usage and `GET _xpack` APIs will match the plugin
name.
This change also removes `IndexLifcycle.NAME` since this was only used
to name the scheduler job and that can be done using
`XPackField.INDEX_LIFECYCLE` instead
Changes to the IndexLifecycleService were necessary since relying on
ClusterChangedEvents for a full picture of the cluster state's settings was
a mistake. It is not necessary that these events hold all settings, especially ones
that are set at node start-up.
Changes to main include:
- move poll interval updates to a SettingsUpdateConsumer
- move scheduler start/stop to a localMasterNodeListener
- keep triggerPolicies in clusterChanged
Changes to tests include:
- removal of some low-level state transition checks in the Service that no longer make sense
since the changes are unconditionally specified in the appropriate listeners
- add integration tests for poll-interval updates
- add integration test assertions for verifying scheduler is started up correctly
* Remove canSetPolicy, canUpdatePolicy and canRemovePolicy
Since we now store a pre-compiled list of steps for an index's phase in the
`PolicyStepsRegistry`, we no longer need to worry about updating policies as any
updates won't affect the current phase, and will only be picked up on phase
transitions.
This also removes the tests that test these methods
Relates to #29823
This is needed as with recent changes to master (see #32952), protocol
is no longer accessible from core, so these classes need to be
duplicated in both places.
This change fixes the shrink action so when the shrink is performed we
remove the single node allocation fromt eh shard allocation filtering
settings. Without this fix replicas cannot be allocated after we have
performed the shrink and we cannot make progress with the rest of the
shink aciton.
This change also fixes a bug in the explain API where the maste node
timeout was being set to null if it wasn't provided instead of using
its default value causing a NPE
If we are waiting on a condition to be met, and the reason
it is not completed is unchanged, we find ourselves updating
cluster state over and over again and kicking of the ILM listeners
to re-check. This is overkill and can generate way too many
cluster state updates
This is the final PR for copying over the necessary components for
clients to parse/render LifecyclePolicy. Changes include:
- move of named-x-content server objects away from client
- move validation into the client copy of LifecyclePolicy
- move LifecycleAction into an interface with `getName`
* Store phase steps for index in PolicyStepsRegistry
This changes the way that steps are retrieved from `PolicyStepsRegistry` to
store the steps on a per-index basis (in memory for now, though that will change
in subsequent PRs). These steps are rebuilt as the index changes phases.
This also fixes a bug where an action with the same phase and name was not being
considered changed (and thus updated) in the compiled steps list. These are now
correctly considered as "upsert" diffs.
Relates to #29823
As we migrate to a per-phase execution model, we need to prepare our cluster-state-step execution model to be aligned. It is the case that the final iteration into the next "currentStep" from the next phase would not be available in the registry yet. This change exits the execution loop early as to not jump into executing the next phase's steps before the registry is properly updated
It is useful to have a random TimeseriesLifecycleType-backed LifecyclePolicy
for testing. This PR exposes a helper method to create one and use it for serialization tests
in LifecyclePolicyTests
The qa tests with security haven't actually gone as far as testing security roles yet, so this is a start in the hopes of both bringing the tests into the ilm plugin
* Adds REST client support for PutOperationMode in ILM
* Corrects licence headers
* iter
* add request converter test
* Fixes tests
* Creates start and stop actions for controlling ILM operation
* Addresses review comments
Since replica counts and allocation rules are set separately, it is not always clear how many replicas are to be allocated in the allocate action. Moving the replicas action to occur at the same time as the allocate action, resolves this confusion that could end an undesired state. This means that the ReplicasAction is removed, and a new optional replicas parameter is added to AllocateAction.
This was originally set to a few seconds while prototyping things.
This interval is for the scheduled trigger of policies. Policies
have this extra trigger beyond just on cluster-state changes because
cluster-state changes may not be happeneing in a cluster for
whatever reason, and we need to continue making progress. Updating
this value to be larger is reasonable since not all operations
are expected to be completed in the span of seconds, but instead in
minutes and hours. 10 minutes is sane.
This commit does the following:
- renames index-lifecycle plugin to ilm
- modifies the endpoints to ilm instead of index_lifecycle
- drops _xpack from the endpoints
- drops a few duplicate endpoints
Gordon Brown
Lee Hinman
Jay Modi
Jason Tedor
Luca Cavanna
Julie Tibshirani
Tal Levy
Tim Brooks
Ryan Ernst
Martijn van Groningen
Jake Landis
Alexander Reelsen
Alpar Torok
Nik Everett
Colin Goodheart-Smithe
Jack Conradson
Lee Hinman