This commit changes the default out-of-the-box configuration for the
number of shards from five to one. We think this will help address a
common problem of oversharding. For users with time-based indices that
need a different default, this can be managed with index templates. For
users with non-time-based indices that find they need to re-shard with
the split API in place they no longer need to resort only to
reindexing.
Since this has the impact of changing the default number of shards used
in REST tests, we want to ensure that we still have coverage for issues
that could arise from multiple shards. As such, we randomize (rarely)
the default number of shards in REST tests to two. This is managed via a
global index template. However, some tests check the templates that are
in the cluster state during the test. Since this template is randomly
there, we need a way for tests to skip adding the template used to set
the number of shards to two. For this we add the default_shards feature
skip. To avoid having to write our docs in a complicated way because
sometimes they might be behind one shard, and sometimes they might be
behind two shards we apply the default_shards feature skip to all docs
tests. That is, these tests will always run with the default number of
shards (one).
The following tokenizers were moved: classic, edge_ngram,
letter, lowercase, ngram, path_hierarchy, pattern, thai, uax_url_email and
whitespace.
Left keyword tokenizer factory in server module, because
normalizers directly depend on it.This should be addressed on a
follow up change.
Relates to #23658
There's no need for an extra `blobExists()` call when writing a blob to the HDFS service. The writeBlob implementation for the HDFS repository already uses the `CreateFlag.CREATE` option on the file creation, which ensures that the blob that's uploaded does not already exist. This saves one network roundtrip.
The TokenMetaData equals method compared byte arrays using `.equals` on
the arrays themselves, which is the equivalent of an `==` check. This
means that a seperate byte[] with the same contents would not be
considered equivalent to the existing one, even though it should be.
The method has been updated to use `Array#equals` and similarly the
hashcode method has been updated to call `Arrays#hashCode` instead of
calling hashcode on the array itself.
This commit removes a test that we can not restore from 1.x and 2.x
repository files. This test is not needed, the version of Elasticsearch
that this commit targets can not even read index files from those
versions.
We have a pile of documentation describing how to rebuild the built in
language analyzers and, previously, our documentation testing framework
made sure that the examples successfully built *an* analyzer but they
didn't assert that the analyzer built by the documentation matches the
built in anlayzer. Unsuprisingly, some of the examples aren't quite
right.
This adds a mechanism that tests that the analyzers built by the docs.
The mechanism is fairly simple and brutal but it seems to be working:
build a hundred random unicode sequences and send them through the
`_analyze` API with the rebuilt analyzer and then again through the
built in analyzer. Then make sure both APIs return the same results.
Each of these calls to `_anlayze` takes about 20ms on my laptop which
seems fine.
When deleting or creating a snapshot for a given shard, elasticsearch
usually starts by listing all the existing snapshotted files in the repository.
Then it computes a diff and deletes the snapshotted files that are not
needed anymore. During this deletion, an exception is thrown if the file
to be deleted does not exist anymore.
This behavior is challenging with cloud based repository implementations
like S3 where a file that has been deleted can still appear in the bucket for
few seconds/minutes (because the deletion can take some time to be fully
replicated on S3). If the deleted file appears in the listing of files, then the
following deletion will fail with a NoSuchFileException and the snapshot
will be partially created/deleted.
This pull request makes the deletion of these files a bit less strict, ie not
failing if the file we want to delete does not exist anymore. It introduces a
new BlobContainer.deleteIgnoringIfNotExists() method that can be used
at some specific places where not failing when deleting a file is
considered harmless.
Closes#28322
Upgrade to lucene-7.4.0-snapshot-1ed95c097b
This version contains:
* An Analyzer for Korean
* An IntervalQuery and IntervalsSource that retrieve minimum intervals of positional queries.
* A new API to retrieve matches (offsets and positions) of a query for a single document.
* Support for soft deletes in the index writer.
* A fixed shingle filter that handles index time synonyms.
* Support for emoji sequence in ICUTokenizer (with an upgrade to icu 61.1)
Fix NPE when CumulativeSum agg encounters null/empty bucket
If the cusum agg encounters a null value, it's because the value is
missing (like the first value from a derivative agg), the path is
not valid, or the bucket in the path was empty.
Previously cusum would just explode on the null, but this changes it
so we only increment the sum if the value is non-null and finite.
This is safe because even if the cusum encounters all null or empty
buckets, the cumulative sum is still zero (like how the sum agg returns
zero even if all the docs were missing values)
I went ahead and tweaked AggregatorTestCase to allow testing pipelines,
so that I could delete the IT test and reimplement it as AggTests.
Closes#27544
This commit removes the http.enabled setting. While all real nodes (started with bin/elasticsearch) will always have an http binding, there are many tests that rely on the quickness of not actually needing to bind to 2 ports. For this case, the MockHttpTransport.TestPlugin provides a dummy http transport implementation which is used by default in ESIntegTestCase.
closes#12792
This commit refactors VersionUtils.resolveReleasedVersions to be
simpler, and in the process fixes the behavior to match that of
VersionCollection.groovy.
closes#30133
Starting with the refactoring in https://github.com/elastic/elasticsearch/pull/22778 (released in 5.3) we may fail to properly replicate operation when a mapping update on master fails. If a bulk
operations needs a mapping update half way, it will send a request to the master before continuing
to index the operations. If that request times out or isn't acked (i.e., even one node in the cluster
didn't process it within 30s), we end up throwing the exception and aborting the entire bulk. This is
a problem because all operations that were processed so far are not replicated any more to the
replicas. Although these operations were never "acked" to the user (we threw an error) it cause the
local checkpoint on the replicas to lag (on 6.x) and the primary and replica to diverge.
This PR does a couple of things:
1) Most importantly, treat *any* mapping update failure as a document level failure, meaning only
the relevant indexing operation will fail.
2) Removes the mapping update callbacks from `IndexShard.applyIndexOperationOnPrimary` and
similar methods for simpler execution. We don't use exceptions any more when a mapping
update was successful.
I think we need to do more work here (the fact that a single slow node can prevent those mappings
updates from being acked and thus fail operations is bad), but I want to keep this as small as I can
(it is already too big).
`javadoc` will switch from detaulting to html4 to html5 in "a future
release". We should get ahead of it so we're not surprised. Also, HTML5
is the future! Er, the present. Anyway, this follows up from #30220 to
make the Javadoc for two of the four remaining projects HTML5
compatible.
The internal test cluster can sometimes have 0 nodes. In this situation,
the http.enabled flag will never be read, and thus no deprecation
warning will be emitted. This commit guards the deprecation warning
check in this case.
Today the translog of an engine is exposed and can be accessed directly.
While this exposure offers much flexibility, it also causes these troubles:
- Inconsistent behavior between translog method and engine method.
For example, rolling a translog generation via an engine also trims
unreferenced files, but translog's method does not.
- An engine does not get notified when critical errors happen in translog
as the access is direct.
This change isolates translog of an engine and enforces all accesses to
translog via the engine.
The index thread pool is no longer needed as its primary use-case for
single-document indexing requests has been relieved now that
single-document indexing requests are converted to bulk indexing
requests (with a single document payload).
This adds 2 testcases that test if a shard goes idle
pending (uncommitted) segments are committed and unreferenced
files will be freed.
Relates to #29482
This change adds the current primary term to the header of the current
translog file. Having a term in a translog header is a prerequisite step
that allows us to trim translog operations given the max valid seq# for
that term.
This commit also updates tests to conform the primary term invariant
which guarantees that all translog operations in a translog file have
its terms at most the term stored in the translog header.
* Add a helper method to get a random java.util.TimeZone
This adds a helper method to ESTestCase that returns a randomized
`java.util.TimeZone`. This can be used when transitioning code from Joda to the
JDK's time classes.
Some features have been deprecated since `6.0` like the `_parent` field or the
ability to have multiple types per index. This allows to remove quite some
code, which in-turn will hopefully make it easier to proceed with the removal
of types.
Currently rest-based tests do not work from the IDE, as the security
manager is configured to permit certain network operations when
using the snapshot jars compiled by gradle. We have an existing
workaround that explicitly associates a codebase with the path
from which the classes are loaded (in this case, the IDE build
directory). This PR adds the rest client to this workaround list.
* Move Streams.copy into elasticsearch-core and make a multi-release jar
This moves the method `Streams.copy(InputStream in, OutputStream out)` into the
`elasticsearch-core` project (inside the `o.e.core.internal.io` package). It
also makes this class into a multi-release class where the Java 9 equivalent
uses `InputStream#transferTo`.
This is a followup from
https://github.com/elastic/elasticsearch/pull/29300#discussion_r178147495
* Move ObjectParser into the x-content lib
This moves `ObjectParser`, `AbstractObjectParser`, and
`ConstructingObjectParser` into the libs/x-content dependency. This decoupling
allows them to be used for parsing for projects that don't want to depend on the
entire Elasticsearch jar.
Relates to #28504
* Fixes query_string query equals timezone check
This change fixes a bug where two `QueryStringQueryBuilder`s were found
to be equal if they had the same timezone set even if the query string
in the builders were different
Closes#29403
* Adds mutate function to QueryStringQueryBuilderTests
* iter
This improves the way similarities are plugged in in order to:
- reject the classic similarity on 7.x indices and emit a deprecation
warning otherwise
- reject unkwown parameters on 7.x indices and emit a deprecation
warning otherwise
Even though this breaks the plugin API, I'd like to backport to 7.x so
that users can get deprecation warnings when they are doing something
that will become unsupported in the future.
Closes#23208Closes#29035
* Begin moving XContent to a separate lib/artifact
This commit moves a large portion of the XContent code from the `server` project
to the `libs/xcontent` project. For the pieces that have been moved, some
helpers have been duplicated to allow them to be decoupled from ES helper
classes. In addition, `Booleans` and `CheckedFunction` have been moved to the
`elasticsearch-core` project.
This decoupling is a move so that we can eventually make things like the
high-level REST client not rely on the entire ES jar, only the parts it needs.
There are some pieces that are still not decoupled, in particular some of the
XContent tests still remain in the server project, this is because they test a
large portion of the pluggable xcontent pieces through
`XContentElasticsearchException`. They may be decoupled in future work.
Additionally, there may be more piecese that we want to move to the xcontent lib
in the future that are not part of this PR, this is a starting point.
Relates to #28504
Removes a set of assertions in the test framework that verified that
Streamable objects could be serialized and deserialized across different
versions. When this was discussed the consensus was that this approach
has not caught many bugs in a long time and that serialization testing of
objects was best left to their respective unit and integration tests.
This commit also removes a transport interceptor that was used in
ESIntegTestCase tests to make these assertions about objects coming in
or off the wire.
Today we have a few problems with how we handle bad requests:
- handling requests with bad encoding
- handling requests with invalid value for filter_path/pretty/human
- handling requests with a garbage Content-Type header
There are two problems:
- in every case, we give an empty response to the client
- in most cases, we leak the byte buffer backing the request!
These problems are caused by a broader problem: poor handling preparing
the request for handling, or the channel to write to when the response
is ready. This commit addresses these issues by taking a unified
approach to all of them that ensures that:
- we respond to the client with the exception that blew us up
- we do not leak the byte buffer backing the request
We historically removed reading from the transaction log to get consistent
results from _GET calls. There was also the motivation that the read-modify-update
principle we apply should not be hidden from the user. We still agree on the fact
that we should not hide these aspects but the impact on updates is quite significant
especially if the same documents is updated before it's written to disk and made serachable.
This change adds back the ability to read from the transaction log but only for update calls.
Calls to the _GET API will always do a refresh if necessary to return consistent results ie.
if stored fields or DocValues Fields are requested.
Closes#26802
Once a document is deleted and Lucene is refreshed, we will not be able
to look up the `version/seq#` associated with that delete in Lucene. As
conflicting operations can still be indexed, we need another mechanism
to remember these deletes. Therefore deletes should still be stored in
the Version Map, even after Lucene is refreshed. Obviously, we can't
remember all deletes forever so a trimming mechanism is needed.
Currently, we remember deletes for at least 1 minute (the default GC
deletes cycle) and clean them periodically. This is, at the moment, the
best we can do on the primary for user facing APIs but this arbitrary
time limit is problematic for replicas. Furthermore, we can't rely on
the primary and replicas doing the trimming in a synchronized manner,
and failing to do so results in the replica and primary making different
decisions.
The following scenario can cause inconsistency between
primary and replica.
1. Primary index doc (index, id=1, v2)
2. Network packet issue causes index operation to back off and wait
3. Primary deletes doc (delete, id=1, v3)
4. Replica processes delete (delete, id=1, v3)
5. 1+ minute passes (GC deletes runs replica)
6. Indexing op is finally sent to the replica which no processes it
because it forgot about the delete.
We can reply on sequence-numbers to prevent this issue. If we prune only
deletes whose seqno at most the local checkpoint, a replica will
correctly remember what it needs. The correctness is explained as
follows:
Suppose o1 and o2 are two operations on the same document with seq#(o1)
< seq#(o2), and o2 arrives before o1 on the replica. o2 is processed
normally since it arrives first; when o1 arrives it should be discarded:
1. If seq#(o1) <= LCP, then it will be not be added to Lucene, as it was
already previously added.
2. If seq#(o1) > LCP, then it depends on the nature of o2:
- If o2 is a delete then its seq# is recorded in the VersionMap,
since seq#(o2) > seq#(o1) > LCP, so a lookup can find it and
determine that o1 is stale.
- If o2 is an indexing then its seq# is either in Lucene (if
refreshed) or the VersionMap (if not refreshed yet), so a
real-time lookup can find it and determine that o1 is stale.
In this PR, we prefer to deploy a single trimming strategy, which
satisfies both requirements, on primary and replicas because:
- It's simpler - no need to distinguish if an engine is running at
primary mode or replica mode or being promoted.
- If a replica subsequently is promoted, user experience is fully
maintained as that replica remembers deletes for the last GC cycle.
However, the version map may consume less memory if we deploy two
different trimming strategies for primary and replicas.
#28245 has introduced the utility class`EngineDiskUtils` with a set of methods to prepare/change
translog and lucene commit points. That util class bundled everything that's needed to create and
empty shard, bootstrap a shard from a lucene index that was just restored etc.
In order to safely do these manipulations, the util methods acquired the IndexWriter's lock. That
would sometime fail due to concurrent shard store fetching or other short activities that require the
files not to be changed while they read from them.
Since there is no way to wait on the index writer lock, the `Store` class has other locks to make
sure that once we try to acquire the IW lock, it will succeed. To side step this waiting problem, this
PR folds `EngineDiskUtils` into `Store`. Sadly this comes with a price - the store class doesn't and
shouldn't know about the translog. As such the logic is slightly less tight and callers have to do the
translog manipulations on their own.
This change refactors the composite aggregation to add an execution mode that visits documents in the order of the values
present in the leading source of the composite definition. This mode does not need to visit all documents since it can early terminate
the collection when the leading source value is greater than the lowest value in the queue.
Instead of collecting the documents in the order of their doc_id, this mode uses the inverted lists (or the bkd tree for numerics) to collect documents
in the order of the values present in the leading source.
For instance the following aggregation:
```
"composite" : {
"sources" : [
{ "value1": { "terms" : { "field": "timestamp", "order": "asc" } } }
],
"size": 10
}
```
... can use the field `timestamp` to collect the documents with the 10 lowest values for the field instead of visiting all documents.
For composite aggregation with more than one source the execution can early terminate as soon as one of the 10 lowest values produces enough
composite buckets. For instance if visiting the first two lowest timestamp created 10 composite buckets we can early terminate the collection since it
is guaranteed that the third lowest timestamp cannot create a composite key that compares lower than the one already visited.
This mode can execute iff:
* The leading source in the composite definition uses an indexed field of type `date` (works also with `date_histogram` source), `integer`, `long` or `keyword`.
* The query is a match_all query or a range query over the field that is used as the leading source in the composite definition.
* The sort order of the leading source is the natural order (ascending since postings and numerics are sorted in ascending order only).
If these conditions are not met this aggregation visits each document like any other agg.
* Remove BytesArray and BytesReference usage from XContentFactory
This removes the usage of `BytesArray` and `BytesReference` from
`XContentFactory`. Instead, a regular `byte[]` should be passed. To assist with
this a helper has been added to `XContentHelper` that will preserve the offset
and length from the underlying BytesReference.
This is part of ongoing work to separate the XContent parts from ES so they can
be factored into their own jar.
Relates to #28504
`$_path` is used by documentation tests to ignore a value from a
response, for example:
```
[source,js]
----
{
"count": 1,
"datafeeds": [
{
"datafeed_id": "datafeed-total-requests",
"state": "started",
"node": {
...
"attributes": {
"ml.machine_memory": "17179869184",
"ml.max_open_jobs": "20",
"ml.enabled": "true"
}
},
"assignment_explanation": ""
}
]
}
----
// TESTRESPONSE[s/"17179869184"/$body.$_path/]
```
That example shows `17179869184` in the compiled docs but when it runs
the tests generated by that doc it ignores `17179869184` and asserts
instead that there is a value in that field. This is required because we
can't predict things like "how many milliseconds will this take?" and
"how much memory will this take?".
Before this change it was impossible to use `$_path` when any component
of the path contained a `.`. This fixes the `$_path` evaluator to
properly escape `.`.
Closes#28770
Changes made in #28972 seems to have changed some assumptions about how
SMILE and CBOR write byte[] values and how this is tested. This changes
the generation of the randomized DocumentField values back to BytesArray
while expecting the JSON and YAML deserialisation to produce Base64
encoded strings and SMILE and CBOR to parse back BytesArray instances.
Closes#29080
Currently we have a fairly complicated logic in the engine constructor logic to deal with all the
various ways we want to mutate the lucene index and translog we're opening.
We can:
1) Create an empty index
2) Use the lucene but create a new translog
3) Use both
4) Force a new history uuid in all cases.
This leads complicated code flows which makes it harder and harder to make sure we cover all the
corner cases. This PR tries to take another approach. Constructing an InternalEngine always opens
things as they are and all needed modifications are done by static methods directly on the
directory, one at a time.
* Decouple XContentBuilder from BytesReference
This commit removes all mentions of `BytesReference` from `XContentBuilder`.
This is needed so that we can completely decouple the XContent code and move it
into its own dependency.
While this change appears large, it is due to two main changes, moving
`.bytes()` and `.string()` out of XContentBuilder itself into static methods
`BytesReference.bytes` and `Strings.toString` respectively. The rest of the
change is code reacting to these changes (the majority of it in tests).
Relates to #28504
As we have factored Elasticsearch into smaller libraries, we have ended
up in a situation that some of the dependencies of Elasticsearch are not
available to code that depends on these smaller libraries but not server
Elasticsearch. This is a good thing, this was one of the goals of
separating Elasticsearch into smaller libraries, to shed some of the
dependencies from other components of the system. However, this now
means that simple utility methods from Lucene that we rely on are no
longer available everywhere. This commit copies IOUtils (with some small
formatting changes for our codebase) into the fold so that other
components of the system can rely on these methods where they no longer
depend on Lucene.
I have long wanted an actual test that dying with dignity works. It is
tricky because if dying with dignity works, it means the test JVM dies
which is usually an abnormal condition. And anyway, how does one force a
fatal error to be thrown. I was motivated to investigate this again by
the fact that I missed a backport to one branch leading to an issue
where Elasticsearch would not successfully die with dignity. And now we
have a solution: we install a plugin that throws an out of memory error
when it receives a request. We hack the standalone test infrastructure
to prevent this from failing the test. To do this, we bypass the
security manager and remove the PID file for the node; this tricks the
test infrastructure into thinking that it does not need to stop the
node. We also bypass seccomp so that we can fork jps to make sure that
Elasticsearch really died. And to be extra paranoid, we parse the logs
of the dead Elasticsearch process to make sure it died with
dignity. Never forget.
Jason Tedor
Martijn van Groningen
Yannick Welsch
Jay Modi
Nik Everett
Stéphane Campinas
Tanguy Leroux
Jim Ferenczi
Zachary Tong
Ryan Ernst
Adrien Grand
Boaz Leskes
Julie Tibshirani
Nhat Nguyen
olcbean
Simon Willnauer
Lee Hinman
tomcallahan
Colin Goodheart-Smithe
Mayya Sharipova
Andy Bristol
Christoph Büscher
Luca Cavanna