This PR merges the `vectors-optimize-brute-force` feature branch, which makes
the following changes to how vector functions are computed:
* Precompute the L2 norm of each vector at indexing time. (#45390)
* Switch to ByteBuffer for vector encoding. (#45936)
* Decode vectors and while computing the vector function. (#46103)
* Use an array instead of a List for the query vector. (#46155)
* Precompute the normalized query vector when using cosine similarity. (#46190)
Co-authored-by: Mayya Sharipova <mayya.sharipova@elastic.co>
Unfortunately, with this change, we won't clean up all unreferenced
generations when reopening. We assume that there's at most one
unreferenced generation when reopening translog. The previous
implementation guarantees this assumption by syncing translog every time
after we remove a translog reader. This change, however, only syncs
translog once after we have removed all unreferenced readers (can be
more than one) and breaks the assumption.
Closes#46267
This reverts commit fd8183ee51d7cf08d9def58a2ae027714beb60de.
Previously we only turned on tests if we saw either `// CONSOLE` or
`// TEST`. These magic comments are difficult for the docs build to deal
with so it has moved away from using them where possible. We should
catch up. This adds another trigger to enable testing: marking a snippet
with the `console` language. It looks like this:
```
[source,console]
----
GET /
----
```
This saves a line which is nice, I guess. But it is more important to me
that this is consistent with the way the docs build works now.
Similarly this enables response testing when you mark a snippet with the
language `console-result`. That looks like:
```
[source,console-result]
----
{
"result": "0.1"
}
----
```
`// TESTRESPONSE` is still available for situations like `// TEST`: when
the response isn't *in* the console-result language (like `_cat`) or
when you want to perform substitutions on the generated test.
Should unblock #46159.
This commit is a follow-up to a change that fixed that multi-get was not
triggering a shard to become search active. In that change, we added a
test that multi-get properly triggers a shard to become search
active. This commit is a follow-up to that change which adds a test for
the get case. While get is already handled correctly in production code,
there was not a test for it. This commit adds one. Additionally, we
factor all the search idle tests from IndexShardIT into a separate test
class, as an effort to keep related tests together instead of a single
large test class containing a jumble of tests, and also to keep test
classes smaller for better parallelization.
Avoid pre-allocating ((N * N) - N) / 2 “BitsIntersector” objects given N filters.
Most adjacency matrices will be sparse and we typically don’t need to allocate all of these objects - can save a lot of allocations when the number of filters is high.
Closes#46212
This commit updates the docs about translog retention and flushing to reflect
recent changes in how peer recoveries work. It also adds some docs to describe
how history is retained for replay using soft deletes and shard history
retention leases.
Relates #45473
Previously, we send recovery requests using CancellableThreads because
we send requests and wait for responses in a blocking manner. With async
recovery, we no longer need to do so. Moreover, if we fail to submit a
request, then we can release the Store using an interruptible thread
which can risk invalidating the node lock.
This PR is the first step to avoid forking when releasing the Store.
Relates #45409
Relates #46178
An exception from the DefaultSearchContext constructor could leak a
searcher, causing future issues like shard lock obtained exceptions. The
underlying cause of the exception in the constructor has been fixed, but
as a safety precaution we also fix the exception handling in
createContext.
Closes#45378
As per #45852 comment we no longer need to log stack-traces in
SecurityTransportExceptionHandler and SecurityHttpExceptionHandler even
if trace logging is enabled.
(cherry picked from commit c99224a32d26db985053b7b36e2049036e438f97)
The test seems to have been failing due to a race condition between
stopping the task and refreshing the destination index. In particular,
we were going forward with refreshing the destination index even
though the task stopped in the meantime. This was fixed in
request.
Closes#43960
Backport of #46271
Previously, when the condition (1st argument) of the IIF function could
be evaluated (folded) to false, the `IfConditional` was eliminated which
caused `IndexOutOfBoundsException` to be thrown when `info()` and
`resolveType()` methods where called.
Fixes: #46268
(cherry picked from commit 9a885a3ac47bc8f52c07770d1d8d670ce0af1e59)
If a primary as being relocated, then the global checkpoint and
retention lease background sync can emit unnecessary warning logs.
This side effect was introduced in #42241.
Relates #40800
Relates #42241
We hit a bug where we can't partially update documents created in a
mixed cluster between 5.x and 6.x. Although this bug does not affect
7.0 or later, we should have a good test that catches this issue.
Relates #46198
When a shard has fallen search idle, and a non-realtime multi-get
request is executed, today such requests do not wait for the shard to
become search active and therefore such requests do not wait for a
refresh to see the latest changes to the index. This also prevents such
requests from triggering the shard as non-search idle, influencing the
behavior of scheduled refreshes. This commit addresses this by attaching
a listener to the shard search active state for multi-get requests. In
this way, when the next scheduled refresh is executed, the multi-get
request will then proceed.
* [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
testSyncFailsIfOperationIsInFlight could fail due to the index request
spawing a GCP sync (new since 7.4). Test now waits for it to finish
before testing that flushed sync fails.
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.
With this change, we can avoid blocking writing threads when trimming
unreferenced readers; hence improving the translog writing performance
in async durability mode.
Close#46201
* Wait for all Rec. to Stop on Node Close
* This issue is in the `RecoverySourceHandler#acquireStore`. If we submit the store release to the generic threadpool while it is getting shut down we never complete the futue we wait on (in the generic pool as well) and fail to ever release the store potentially.
* Fixed by waiting for all recoveries to end on node close so that we aways have a healthy thread pool here
* Closes#45956
When some high values are randomly picked up - for example the number
of indices to snapshot or the number of snapshots to create - the tests
in S3BlobStoreRepositoryTests can generate a high number of requests to
the internal S3 server.
In order to test the retry logic of the S3 client, the internal server is
designed to randomly generate random server errors. When many
requests are made, it is possible that the S3 client reaches its maximum
number of successive retries capacity. Then the S3 client will stop
retrying requests until enough retry attempts succeed, but it means
that any request could fail before reaching the max retries count and
make the test fail too.
Closes#46217Closes#46218Closes#46219
Fix test issue to stabilise scoring through use of DFS search mode.
Randomised index-then-delete docs introduced by the test framework likely caused an imbalance in IDF scores across shards. Also made number of shards used in test a random number for added test coverage.
Closes#46174
Previously, if the DataType of all the WHEN conditions of a CASE
statement is NULL, then it was set to NULL even if the ELSE clause
has a non-NULL data type, e.g.:
```
CASE WHEN a = 1 THEN NULL
WHEN a = 5 THEN NULL
ELSE 'foo'
```
Fixes: #46032
(cherry picked from commit 8c1012efbbd3a300afd0dfb9b18250f15ea753f9)
This is the first PR that merges changes made to server module from
the enrich branch (see #32789) into the master branch.
The plan is to merge changes made to the server module separately from
the pr that will merge enrich into master, so that these changes can
be reviewed in isolation.
Some netty behavior is controlled by system properties. While we want to
test with the defaults for Elasticsearch for most tests, within netty we
want to ensure these netty settings exhibit correct behavior. This
commit adds variants of test and integTest tasks for netty which set the
unpooled and direct buffer pooled allocators.
relates #45881
Though we allow CCS within datafeeds, users could prevent nodes from accessing remote clusters. This can cause mysterious errors and difficult to troubleshoot.
This commit adds a check to verify that `cluster.remote.connect` is enabled on the current node when a datafeed is configured with a remote index pattern.
Julie Tibshirani
Nhat Nguyen
Nik Everett
Aleh Zasypkin
James Rodewig
Lisa Cawley
Jason Tedor
markharwood
David Turner
Albert Zaharovits
István Zoltán Szabó
Henning Andersen
Alpar Torok
Andrey Ershov
Dimitris Athanasiou
Marios Trivyzas
Lee Hinman
Benjamin Trent
Zachary Tong
David Roberts
Ignacio Vera
dengweisysu
Anup
Armin Braun
Tanguy Leroux
Ioannis Kakavas
Martijn van Groningen
Jay Greenberg
Ryan Ernst