This PR extends the TranslogDeletionPolicy to allow keeping the translog files longer than what is needed for recovery from lucene. Specifically, we allow specifying the total size of the files and their maximum age (i.e., keep up to 512MB but no longer than 12 hours). This will allow making ops based recoveries more common.
Note that the default size and age still set to 0, maintaining current behavior. This is needed as the other components in the system are not yet ready for a longer translog retention. I will adapt those in follow up PRs.
Relates to #10708
This commit does two things:
1. Adds logging at the DEBUG level for when the index-N blob is
updated.
2. When attempting to delete a snapshot, if the snapshot was not found
in the repository data, an exception is now thrown instead of silently
ignoring the lack of presence of the snapshot in the repository data.
We use assertBusy in many places where the underlying code throw exceptions. Currently we need to wrap those exceptions in a RuntimeException which is ugly.
This commit adds a NamedXContentProvider interface that can
be implemented by plugins or modules using Java's SPI feature
in order to provide additional NamedXContent parsers to external
applications like the Java High Level Rest Client.
At index time Elasticsearch needs to look up the version associated with the
`_id` of the document that is being indexed, which is often the bottleneck for
indexing.
While reviewing the output of the `jfr` telemetry from a Rally benchmark, I saw
that significant time was spent in `ConcurrentHashMap#get` and `ThreadLocal#get`.
The reason is that we cache lookup objects per thread and segment, and for every
indexed document, we first need to look up the cache associated with this
segment (`ConcurrentHashMap#get`) and then get a state that is local to the
current thread (`ThreadLocal#get`). So if you are indexing N documents per
second and have S segments, both these methods will be called N*S times per
second.
This commit changes version lookup to use a cache per index reader rather than
per segment. While this makes cache entries live for less long, we now only need
to do one call to `ConcurrentHashMap#get` and `ThreadLocal#get` per indexed
document.
This snapshot has faster range queries on range fields (LUCENE-7828), more
accurate norms (LUCENE-7730) and the ability to use fake term frequencies
(LUCENE-7854).
This commit renames the needsScores method so as to make it
automatically generatable, based on the name of the `_score` variable
which is available in search scripts. It also adds documentation to
ScriptContext to explain the naming and signature of such methods.
This commit removes the global caching of the field query and replaces it with
a caching per field. Each field can use a different `highlight_query` and the rewriting of
some queries (prefix, automaton, ...) depends on the targeted field so the query used for highlighting
must be unique per field.
There might be a small performance penalty when highlighting multiple fields since the query needs to be rewritten
once per highlighted field with this change.
Fixes#25171
* Remove QUERY_AND_FETCH BWC for pre-5.3.0 nodes
This was a BWC layer where we expicitly set the `search_type` to
"query_and_fetch" when a single node is queried on pre-5.3 nodes. Since 6.0 no
longer needs to be compatible with 5.3 nodes, this can be removed.
* Fix indentation
* Remove unused QUERY_FETCH_ACTION_NAME constant
* Add more missing AggregationBuilder getters
- getMetadata for all aggs
- various getters on TermsAggBuilder (without "get" prefix to maintain convention)
- Also makes InternalSum's ctor public, to follow suit of other metrics (min/max/avg/etc)
We introduced a new API for ranges in order to be able to decide whether points
or doc values would be more appropriate to execute a query, but since
`ProfileWeight` does not implement this API, the optimization is disabled when
profiling is enabled.
In order to add scroll support for cross cluster search we need
to resolve the nodes encoded in the scroll ID to send requests to the
corresponding nodes. This change adds the low level connection infrastructure
that also ensures that connections are re-established if the cluster is
disconnected due to a network failure or restarts.
Relates to #25094
Today if a channel gets closed due to a disconnect we notify the response
handler that the connection is closed and the node is disconnected. Unfortunately
this is not a complete solution since it only works for published connections.
Connections that are unpublished ie. for discovery can indefinitely hang since we
never invoke their handers when we get a failure while a user is waiting for
the response. This change adds connection tracking to TcpTransport that ensures
we are notifying the corresponding connection if there is a failure on a channel.
This modifies a method Mark added to the AggregatorBase that allows aggregations
to add additional memory tracking for datastructures used during execution. If
an aggregation would like to reclaim circuit breaker reserved bytes by adding a
negative number, `addWithoutBreaking` should be used instead of
`addEstimateBytesAndMaybeBreak`.
Resolves#24511
Duplicate data paths already fail to work because we would attempt to
take out a node lock on the directory a second time which will fail
after the first lock attempt succeeds. However, how this failure
manifests is not apparent at all and is quite difficult to
debug. Instead, we should explicitly reject duplicate data paths to make
the failure cause more obvious.
Relates #25178
When attempting to obtain the node lock, if an exception is thrown it is
not logged. This makes debugging difficult. This commit causes such an
exception to be logged.
Relates #25176
* Aggregations bug: Significant_text fails on arrays of text.
The set of previously-seen tokens in a doc was allocated per-JSON-field string value rather than once per JSON document meaning the number of docs containing a term could be over-counted leading to exceptions from the checks in significance heuristics. Added unit test for this scenario
Closes#25029
Sorted scroll search can use early termination when the index sort matches the scroll search sort.
The optimization can be done after the first query (which still needs to collect all documents)
by applying a query that only matches documents that are greater than the last doc retrieved in the previous request.
Since the index is sorted, retrieving the list of documents that are greater than the last doc
only requires a binary search on each segment.
This change introduces this new query called `SortedSearchAfterDocQuery` and apply it when possible.
Scrolls with this optimization will search all documents on the first request and then will early terminate each segment
after $size doc for any subsequent requests.
Relates #6720
#25005 changed the translog dynamic to fsync the checkpoint before trimming a file. This changed the dynamics of potential failure modes which requires a change to testWithRandomException - it's now possible that we had an exception but the translog was trimmed.
Closes#25133
Get mappings HEAD requests incorrectly return a content-length header of
0. This commit addresses this by removing the special handling for get
mappings HEAD requests, and just relying on the general mechanism that
exists for handling HEAD requests in the REST layer.
Relates #23192
Today when an exception is thrown handling a HEAD request, the body is
swallowed before the channel has a chance to see it. Yet, the channel is
where we compute the content length that would be returned as a header
in the response. This is a violation of the HTTP specification. This
commit addresses the issue. To address this issue, we remove the special
handling in bytes rest response for HEAD requests when an exception is
thrown. Instead, we let the upstream channel handle the special case, as
we already do today for the non-exceptional case.
Relates #25172
We have a custom logger implementation known as a prefix logger that is
used to write every message by the logger with a given prefix. This is
useful for node-level, index-level, and shard-level messages where we
want to log the node name, index name, and shard ID, respectively, if
possible. The mechanism that we employ is that of a marker. Log4j has a
built-in facility for managing these markers, but its effectively a
memory leak because these markers are held in a map and can never be
released. This is problematic for us since indices and shards do not
necessarily have infinite life spans and so on a node where there are
many indices being creted and destroyed, this infinite lifespan can be a
problem indeed. To solve this, we use our own cache of markers. This is
necessary to prevent too many instances of the marker for the same
prefix from being created (just think of all the shard-level components
that exist in the system), and to workaround the effective leak in
Log4j. These markers are stored as weak references in a weak hash
map. It is these weak references that are unneeded. When a key is
removed from a weak hash map, the corresponding entry is placed on a
reference queue that is eventually cleared. This commit simplifies
prefix logger by removing this unnecessary weak reference wrapper.
Relates #22460
When the cluster state is updated with Shard Started entries, it simply adds "shard-started" as the source of the change.
This adds the index name and shard ID so that we can see who/what is spamming the changes when the index creation step has already left the cluster state.
There are a few places where arrays are output in messages yet the
output would merely use the default toString implementation rather than
actually putting the content of the array in the message. This commit
fixes the issue.
Relates #24340
This change extends the tests and parsing of SearchResponse to make sure we can
skip additional fields the parser doesn't know for forward compatibility
reasons.
This commit adds back "id" as the key within a script to specify a
stored script (which with file scripts now gone is no longer ambiguous).
It also adds "source" as a replacement for "code". This is in an attempt
to normalize how scripts are specified across both put stored scripts and script usages, including search template requests. This also deprecates the old inline/stored keys.
When `index.mapping.single_type` is `true` the `_uid` field is not used and instead `_id` field is used.
Prior to this change nested documents would in this case still use the `_uid` field to mark to what root
document they belong to. In case of deleting documents this could lead to only the root Lucene document
to be deleted and not the nested Lucene documents. This broke the docid block ordering the block join
relies on in order to work correctly and thus causing the `nested` query, `nested` aggregation, nested sorting
and nested inner hits to either fail or yield incorrect results.
This bug only manifests in 6.0.0-ALPHA2 release and snaphots (5.5.0-SNAPSHOT, 5.6.0-SNAPSHOT, 6.0.0-SNAPSHOT).
This was introduced in #24460: the constructor of `Translog.Delete` that takes
a `StreamInput` does not set the type and id. To make it a bit more robust, I
made fields final so that forgetting to set them would make the compiler
complain.
This change removes the `postings` highlighter. This highlighter has been removed from Lucene master (7.x) because it behaves
exactly like the `unified` highlighter when index_options is set to `offsets`:
https://issues.apache.org/jira/browse/LUCENE-7815
It also makes the `unified` highlighter the default choice for highlighting a field (if `type` is not provided).
The strategy used internally by this highlighter remain the same as before, it checks `term_vectors` first, then `postings` and ultimately it re-analyzes the text.
Ultimately it rewrites the docs so that the options that the `unified` highlighter cannot handle are clearly marked as such.
There are few features that the `unified` highlighter is not able to handle which is why the other highlighters (`plain` and `fvh`) are still available.
I'll open separate issues for these features and we'll deprecate the `fvh` and `plain` highlighters when full support for these features have been added to the `unified`.
This change extends the tests and parsing of SearchShardFailure to make sure we
can skip fields the parser doesn't know for forward compatibility reasons.
When parsing responses we should be ignoring any new unknown fields or inner
objects in most cases to be forward compatible with changes in core on the
client side. This change adds test for this for Suggestions and its various
subclasses to check if we are able to ignore new fields and objects in the
xContent.
Boaz Leskes
Ali Beyad
Christoph Büscher
Guillaume Le Floch
Martijn van Groningen
Jim Ferenczi
markharwood
Adrien Grand
Tanguy Leroux
Ryan Ernst
Lee Hinman
Zachary Tong
Lee Hinman
Simon Willnauer
Sergey Galkin
Jason Tedor
Chris Earle
Sergey Novikov
Koen De Groote
Tal Levy