* [ML][Transforms] fixing rolling upgrade continuous transform test (#45823)
* [ML][Transforms] fixing rolling upgrade continuous transform test
* adjusting wait assert logic
* adjusting wait conditions
* [ML][Transforms] allow executor to call start on started task (#46347)
* making sure we only upgrade from 7.4.0 in test
* [ML][Transforms] fixing stop on changes check bug
* Adding new method finishAndCheckState to cover race conditions in early terminations
* changing stopping conditions in `onStart`
* allow indexer to finish when exiting early
Fixes a problem where operations_behind would be one less than
expected per shard in a new index matched by the data frame
transform source pattern.
For example, if a data frame transform had a source of foo*
and a new index foo-new was created with 2 shards and 7 documents
indexed in it then operations_behind would be 5 prior to this
change.
The problem was that an empty index has a global checkpoint
number of -1 and the sequence number of the first document that
is indexed into an index is 0, not 1. This doesn't matter for
indices included in both the last and next checkpoints, as the
off-by-one errors cancelled, but for a new index it affected
the observed result.
Adds index versioning for the internal data frame transform index. Allows for new indices to be created and referenced, `GET` requests now query over the index pattern and takes the latest doc (based on INDEX name).
* [ML][Data frame] fixing failure state transitions and race condition (#45627)
There is a small window for a race condition while we are flagging a task as failed.
Here are the steps where the race condition occurs:
1. A failure occurs
2. Before `AsyncTwoPhaseIndexer` calls the `onFailure` handler it does the following:
a. `finishAndSetState()` which sets the IndexerState to STARTED
b. `doSaveState(...)` which attempts to save the current state of the indexer
3. Another trigger is fired BEFORE `onFailure` can fire, but AFTER `finishAndSetState()` occurs.
The trick here is that we will eventually set the indexer to failed, but possibly not before another trigger had the opportunity to fire. This could obviously cause some weird state interactions. To combat this, I have put in some predicates to verify the state before taking actions. This is so if state is indeed marked failed, the "second trigger" stops ASAP.
Additionally, I move the task state checks INTO the `start` and `stop` methods, which will now require a `force` parameter. `start`, `stop`, `trigger` and `markAsFailed` are all `synchronized`. This should gives us some guarantees that one will not switch states out from underneath another.
I also flag the task as `failed` BEFORE we successfully write it to cluster state, this is to allow us to make the task fail more quickly. But, this does add the behavior where the task is "failed" but the cluster state does not indicate as much. Adding the checks in `start` and `stop` will handle this "real state vs cluster state" race condition. This has always been a problem for `_stop` as it is not a master node action and doesn’t always have the latest cluster state.
closes#45609
Relates to #45562
* [ML][Data Frame] moves failure state transition for MT safety (#45676)
* [ML][Data Frame] moves failure state transition for MT safety
* removing unused imports
This commit replaces task_state and indexer_state in the
data frame _stats output with a single top level state
that combines the two. It is defined as:
- failed if what's currently reported as task_state is failed
- stopped if there is no persistent task
- Otherwise what's currently reported as indexer_state
Backport of #45276
* [ML][Data Frame] Add update transform api endpoint (#45154)
This adds the ability to `_update` stored data frame transforms. All mutable fields are applied when the next checkpoint starts. The exception being `description`.
This PR contains all that is necessary for this addition:
* HLRC
* Docs
* Server side
This adds support for `geo_bounds` aggregation inside the `pivot.aggregations` configuration.
The two points returned from the `geo_bounds` aggregation are transformed into `geo_shape` whose types are dynamic given the point's similarity.
* `point` if the two points are identical
* `linestring` if the two points share either a latitude or longitude
* `polygon` if the two points are completely different
The automatically deduced mapping for the resulting field is a `geo_shape`.
introduces an abstraction for how checkpointing and synchronization works, covering
- retrieval of checkpoints
- check for updates
- retrieving stats information
This is a followup to #44350. The indexer stats used to
be persisted standalone, but now are only persisted as
part of a state-and-stats document. During the review
of #44350 it was decided that we'll stick with this
design, so there will never be a need for an indexer
stats object to store its transform ID as it is stored
on the enclosing document. This PR removes the indexer
stats document ID.
Backport of #44768
This change adjusts the data frame transforms stats
endpoint to return a structure that is easier to
understand.
This is a breaking change for clients of the data frame
transforms stats endpoint, but the feature is in beta so
stability is not guaranteed.
Backport of #44350
Since #44344 we use IndicesOptions.LENIENT_EXPAND_OPEN
when deciding which indices to include in checkpoint
calculation. This change uses the same option when
deciding which indices to search for data and which
indices to get mappings from, otherwise there is a
potential mismatch between the checkpoint details and
what is searched elsewhere.
This adds a new dynamic cluster setting `xpack.data_frame.num_transform_failure_retries`.
This setting indicates how many times non-critical failures should be retried before a data frame transform is marked as failed and should stop executing. At the time of this commit; Min: 0, Max: 100, Default: 10
* [ML][Data Frame] treat bulk index failures as an indexing failure
* removing redundant public modifier
* changing to an ElasticsearchException
* fixing redundant public modifier
make checkpointing more robust:
- do not let checkpointing fail if indexes got deleted
- treat missing seqNoStats as just created indices (checkpoint 0)
- loglevel: do not treat failed updated checks as error
fixes#43992
This commit creates new base classes for master node actions whose
response types still implement Streamable. This simplifies both finding
remaining classes to convert, as well as creating new master node
actions that use Writeable for their responses.
relates #34389
Rewrites how continuous data frame transforms calculates and handles buckets that require an update. Instead of storing the whole set in memory, it pages through the updates using a 2nd cursor. This lowers memory consumption and prevents problems with limits at query time (max_terms_count). The list of updates can be re-retrieved in a failure case (#43662)
Previously a data frame transform would check whether the
source index was changed every 10 seconds. Sometimes it
may be desirable for the check to be done less frequently.
This commit increases the default to 60 seconds but also
allows the frequency to be overridden by a setting in the
data frame transform config.
Data frame task responses had logic to return a HTTP 500 status code if there was
any node or task failures even if other tasks in the same request reported correctly.
This is different to how other task responses are handled where a 200 is always
returned leaving the client should check for failures. Returning a 500 also breaks
the high level rest client so always return a 200
Closes#44011
Hendrik Muhs
Benjamin Trent
David Roberts
Przemysław Witek
Armin Braun
Ryan Ernst
David Kyle