Today we maintain a lot of settings on the shard level which are all index level settings.
In order to cut over to the new settings API where we register update listener we have to move
all of them on to the index level otherwise we need a way to un-register listeners which is error-prone
and requires additional handling when shards are closed. It's simpler and also more accurate to handle all of
them on the index level where we can trash the entire registry for update listener once the index goes out of scope.
Don't set the suppressed Exception in Translog.closeOnTragicEvent(Exception ex) if it is an
AlreadyClosedException. ACE is thrown by the TranslogWriter and as cause might
contain the Exception that we add the suppressed ACE to. We then end up with a
circular reference where Exception A has a suppressed Exception B that has as cause A.
This would cause a stackoverflow when we try to serialize it.
For a more detailed description see #15941closes#15941
This commit adds load averages to the OS stats on FreeBSD. For these
stats to be available, linprocfs must be available and mounted at
/compat/linux/proc.
`refresh_interval` is a per index setting but we interpret and maintain it per shard. This
change moves the refresh task outside of IndexShard to the IndexService where it logically belongs
and reuses scheduling infrastructure used for translog fsync (async commit).
This change will use the same task for all shards of an index while previously we used on thread/task
per shard to refresh. This will also prevent too many concurrent refreshes if there are many indices and
shards allocated on a single node.
These filters leak into highlighting and probably other places and cause
things like the type name to be highlighted when using
requireFieldMatch=false. We could have special hacks to keep them out of
highlighting but it feals better to keep them out of any variable named
"originalQuery".
Closes#15689
This commit adds convenience methods to o.e.t.t.CapturingTransport
that enables capturing requests and clearing the captured requests
with a single method. This is to simplify a common pattern in tests of
capturing requests, and then clearing the captured requests.
Also renaming internal methods to reflect that they are dealing with
jts coordinates. Also renamed the list() to build() method for creating
the coordinates lists and adding constructors to PolygonBuilder that
take CoordinatesBuilders and implicitely call build() on them.
The indexing buffer on a node (default: 10% of the JVM heap) is now a "shared pool" across all shards on that node. This way, shards doing intense indexing can use much more than other shards doing only light indexing, and only once the sum of all indexing buffers across all shards exceeds the node's indexing buffer will we ask shards to move recently indexed documents to segments on disk.
determined, the UpdateRequest would still try to parse the content instead
of throwing the standard ElasticsearchParseException. This manifests when
passing illegal JSON in the request body that does not begin with a '{'.
By trying to parse the content from an unknown request body content type,
the UpdateRequest was throwing a null pointer exception. This has been
fixed to throw an ElasticsearchParseException, to be consistent with the
behavior of all other requests in the face of undecipherable request
content types.
Closes#15822
1. Uses forbidden patterns to prevent things from referencing
java.io.Serializable or from mentioning serialVersionUID.
2. Uses -Xlint:-serial so we don't have to hear from javac that we aren't
declaring serialVersionUID on any classes that we make that happen to extend
Serializable.
3. Remove Serializable and serialVersionUID declarations.
I didn't use forbidden apis because it doesn't look like it has a way to ban
explicitly implementing Serializable. If you try to ban Serializable with
forbidden apis you end up banning all Exceptions and all Strings.
Closes#15847
So far the validation of geo shapes was only taking place in the
parse methods in ShapeBuilder. With the recent refactoring we no
longer can rely on shapes being parsed from json, so the same kind
of validation should take place when just using the java api.
A lot of validation concerns the number of points a shape needs to
have in order to be valid. Since this is not possible with current
builders where points can be added one by one, the builder constructors
are changed to require the mandatory parameters and validate those
already at construction time. To help with constructing longer lists
of points, a new utility PointsListBuilder is instroduces which can
produce list of coordinates accepted by most of the other shape builder
constructors.
Also adding tests for invalid shape exceptions to the already existing
shape builder tests.
Running requests via the percolate or mpercolate api is irrelevant.
What is relevant is that when nodes come back that they report the expected number of matches.
We fail today with ClusterBlockExceptions if an alias expands to a closed index
during search since we miss to check the index option down the road after we expanded
aliases.
Closes#13278
This commit removes and now forbids use of
org.apache.lucene.index.IndexWriter#isLocked as this method was
deprecated in LUCENE-6508. The deprecation is due to the fact that
checking if a lock is held before acquiring that lock is subject to a
time-of-check-to-time-of-use race condition. There were three uses of
IndexWriter#isLocked in the code base:
- a logging statement in o.e.i.e.InternalEngine where we are already in
an exceptional condition that the lock was held; in this case,
logging whether or not the directory is locked is superfluous
- in o.e.c.l.u.VersionsTests where we were verifying that a write lock
is released upon closing an IndexWriter; in this case, the check is
not needed as successfully closing an IndexWriter releases its
write lock
- in o.e.t.s.MockFSDirectoryService where we were verifying that a
directory is not write-locked before (implicitly) trying to obtain
such a write lock in org.apache.lucene.index.CheckIndex#<init> (this
is the exact type of a situation that is subject to a race
condition); in this case we can proceed by just (implicitly) trying
to obtain the write lock and failing if we encounter a
LockObtainFailedException
This commit reduces the former ShardIndexinService to a simple stats/metrics
class, moves IndexingSlowLog to the IndexService level since it can be shared
across shards of an index and is now hidden behind IndexingOperationListener.
IndexingOperationListener is now a first class citizen in IndexShard and is passed
in from IndexService.
It had some funky errors, like lenient:true not working and queries with
two integer fields blowing up if there was no analyzer defined on the
query. This throws a bunch more tests at it and rejiggers how non-strings
are handled so they don't wander off into scary QueryBuilder-land unless
they have a nice strong analyzer to protect them.
Closes#15860
It currently tries to align to the page size (16KB) by default. However, this
might waste a significant memory (if many BytesStreamOutputs are allocated)
and is also useless given that BytesStreamOutput does not recycle (on the
contrary to ReleasableBytesStreamOutput). So the initial size has been changed
to 0.
Closes#15789
This commit removes and now forbids all uses of
java.util.concurrent.ThreadLocalRandom across the codebase. The
underlying issue with ThreadLocalRandom is that it can not be
seeded. This means that if ThreadLocalRandom is used in production code,
then tests that cover any code path containing ThreadLocalRandom will be
prevented from being reproducible by use of ThreadLocalRandom. Instead,
using org.elasticsearch.common.random.Randomness#get will give
reproducible sources of random when running under tests and otherwise
still give an instance of ThreadLocalRandom when running as production
code.
On slow machines when this test randomly picks a large number of shards it can occasionally take more than 32.5 seconds to snapshot all shards. That is causing the test to miss the second to last assert in awaitsBusy at 32.5 seconds and then timeout in BlockingClusterStateListener at 60 seconds. Due to the timeout, the pending task queue is cleaned before the last awaitsBusy assert at 65 seconds and as a result the last assert runs on a completely empty queue and fails with a very confusing assert error.
This commit makes the timeout in BlockingClusterStateListener to occur after the last assert in assertBusyPendingTasks and therefore allows assertBusyPendingTasks to perform the last assert before cleaning the pending tasks queue takes place.
This commit also reduces the maximum number of shards used in the test to 10 in order to speed up this test.
We have a bug that makes all per-index bitset caches store bitsets for all
indices. In the case that you have many indices, which is fairly common with
time-based data, this could translate to a lot of wasted memory.
Closes#15820
`ScheduledThreadPoolExecutor` allows you to schedule tasks to run once or periodically at the future. If such a task throws an exception, that exception is caught and reported in the future that `ScheduledThreadPoolExecutor#schedule` returns. However, we typically do not capture the future / do not test it for errors. This results in exception being swallowed and not reported. To mitigate this we now wrap any command in a LoggingRunnable (already used for periodic tasks). Also, RunnableCommand is changed not to swallow exception but percolate them further for reporting by the future.
Closes#15824
This commit adds a new test that can throw an IOException at any point in time
and ensures that all previously synced documents can be successfully recovered after hitting
an excepiton.
Relates to #15788
Warmers are now barely useful and will be removed in 3.0. Note that this only
removes the warmer API and query-based warmers. We still have warmers internally
for eg. global ordinals.
Close#15607
Right now we define the same sort of methods as taking String arrays and
string varargs. We should standardize on one and varargs is easier to
call so lets use varargs!
This commit uses a CyclicBarrier to correctly and simply sychronize the
driver and test threads in
ClusterServiceIT#testClusterStateUpdateTasksAreExecutedInOrder.
This commit speeds up GeoShapeQueryTests by reducing the size of the random generated shapes and defaulting geo_shape indexes to use quadtree (more efficient for shapes) over geohash.
test failed, because now the percolator returns upto 10 matches whereas before this was unbounded. The test has been updated to take this in account by checking the total count instead of the number of matches
This commit cleans up some of the assertions in
TransportReplicationActionTests#runReplicateTest:
- use a Map to track actual vs. expected requests
- assert that no request was sent to the local node
- use RoutingTable#shardRoutingTable convenience method
- explicitly use false in boolean conditions
- clarify requests are expected on replica shards when assigned and
execution on replicas is true
- test ShardRouting equality when checking the failed shard request
This commit adds tighter assertions in
TransportReplicationActionTests#runReplicateTest that replication
requests are sent to the correct shard copies.
This commit addresses an issue when handling a failed replication
request against a relocating target shard. Namely, if a replication
request fails against the target of a relocation we currently fail both
the source and the target. This leads to an unnecessary
recovery. Instead, only the target of the relocation should be failed.
* Added percolator field mapper that extracts the query terms and indexes these terms with the percolator query.
* At percolate time these extracted terms are used to query percolator queries that are like to be evaluated. This can significantly cut down the time it takes to percolate. Whereas before all percolator queries were evaluated if they matches with the document being percolated.
* Changes made to percolator queries are no longer immediately visible, a refresh needs to happen before the changes are visible.
* By default the percolate api only returns upto 10 matches instead of returning all matching percolator queries.
* Made percolate more modular, so that it is easier to add unit tests.
* Added unit tests for the percolator.
Closes#12664Closes#13646
We today delete the translog-N.tlog file if any subsequent operation fails
but we might actually be in a good state if for instance the creation of the writer
failes after we sucessfully baked the new translog generation into the checkpoint. In this situation
we used to delete the translog-N.tlog file and failed on the next recovery of the translog with a
NoSuchFileException | FileNotFoundException just like in https://discuss.elastic.co/t/cannot-recover-index-because-of-missing-tanslog-files/38336
This commit changes the behavior and cleans up that limbo state on recovery if we already have a generation+1 file written but not baked into
the checkpoint we remove that file but only if the previous ckp file has already been renamed otherwise we know we can't be in this state.
We has a postIndex|DeleteUnderLock listener callback to load percolator
queries which is entirely private to the index shard in the meanwhile. Yet,
it still calls an external callback while holding an indexing lock which is scary
since we have no control over how long the operation could possibly take.
This commit decouples the percolator registry entirely from the ShardIndexingService
by pessimistically fetching percolator documents from the the engine using realtime get.
Even in situations where the same document is changed concurrently we will eventually end up
in the correct state without loosing an update. This also moves the index throtteling stats directly into
the engine to entirely remove the need for the dependency between InternalEngine and ShardIndexingService.
Relocating a non-primary shard from one node to another is actually done by recovering from the active
primary shard in the cluster, and not the node that we are logically relocating from.
Closes#15775
This commit addresses an issue where a cluster state task listener
throwing an exception could prevent other listeners from being notified,
and could prevent the executor from receiving notifications that a new
cluster state was published. Additionally, this commit also addresses a
similar issue for executors handling cluster state publication
notifications.
Adds task manager class and enables all activities to register with the task manager. Currently, the immutable Transport*Activity class represents activity itself shared across all requests. This PR adds and an additional structure Task that keeps track of currently running requests and can be used to communicate with these requests using TransportTaskAction.
Related to #15117
We used to write into an in-memory buffer and if necessary also allow reading
from the memory buffer if the some translog locations that are not flushed to
the channel need to be read. This commit hides all writing behind a buffered output
stream and if ncecessary flushes all buffered data to the channel for reading. This allows
for several simplifcations like reusing javas build in BufferedOutputStream and removes the
need for read write locks on the translog writer. All thread safety is now achived using
the synchronized primitive.
Several IOExceptions are always wrapped in an NotSerializableWrapper which is
annoying to read. These exceptions are important to get right across the network
and we should support the important ones that indicate problems on the Filesystem.
This commit also adds general support for IOException to preserve the parent type
across the network if no specific type is serializable.
As a default in V2, the GeoPointField.stored option was set to true. Since this consumes disk space with no positive benefit the default stored option is being reverted back to false.
This commit restores logging the ShardRouting#shardId at the front of
the log messages in ShardStateAction. The reason for this is so that
shard-level log messages have the format "[component][node][shard]
message".
There are two bugs:
- the 'global_ordinals_low_cardinality' mode requires a fielddata-based impl so
that it can extract the segment to global ordinal mapping
- the 'global_ordinals_hash' mode abusively casts to the values source to a
fielddata-based impl while it is not needed
Closes#14882
This commit modifies the handling of cluster states in
o.e.c.a.s.ShardStateAction so that all necessary state is obtained
externally to the ShardStateAction#shardFailed and
ShardStateAction#shardStarted methods. This refactoring permits the
removal of the ClusterService field from ShardStateAction.
This commit applies a minor code cleanup to
o/e/c/ClusterStateObserver.java. In particular
- employ the diamond operator instead of explicitly specifying a
generic type parameter
- use 'L' instead of 'l' for specifying a long literal
- remove redundant static modifier on a nested interface
- remove redundant public access modifiers on interface methods
- reformat the declaration of the four-argument ChangePredicate#apply
- simplify the bodies of ValidationPredicate#apply
This commit fixes multiField support for GeoPointFieldMapper by passing an externalValueContext to the multiField parser. Unit testing is added for multi field coverage.
The MapperService doesn't currently check the
index.mapper.dynamic setting during index creation,
so indices can be created with dynamic mappings even
if this setting is false. Add a check that throws an
exception in this case. Fixes#15381
DedicatedClusterSnapshotRestoreIT#testRestoreIndexWithMissingShards took ~1.5 min to finish
due to timeouts that are applied if not all shards are allocated. Now that the index that has
unallocated shareds is not refreshed the test is more reasonable and runs in 15 sec
With this commit we check more precisely on the result of a bulk
request. It could either be ok, fail or be rejected due to resource
constraints. Previously, we have relied that by default we never
get rejected.
However, this is a valid condition even when retrying. With this
commit we check that we either retried often enough that we don't
get rejected *and* if we got rejected that we maxed out the number
of specified retries.
When specifying a string field, you can either do:
```
{
"foo": "bar"
}
```
or
```
{
"foo": {
"value": "bar",
"boost": 42
}
}
```
The latter option is now removed.
Closes#15388
Removal of the pattern node.addShard() -> calculate weight -> node.removeShard() which is expensive as, beside map lookups, it invalidates caching of precomputed values in ModelNode and ModelIndex. Replaced by adding an additional parameter to the weight function which accounts for the added / removed shard.
This removes the backward compatibility layer with pre-2.0 indices, notably
the extraction of _id, _routing or _timestamp from the source document when a
path is defined.
Today when dynamically mapping a field that is already defined in another type,
we use the regular dynamic mapping logic and try to copy some settings to avoid
introducing conflicts. However this is quite fragile as we don't deal with every
existing setting. This proposes a different approach that will just reuse the
shared field type.
Close#15568
FunctionScoreQuery should do two things that it doesn't do today:
- propagate the two-phase iterator from the wrapped scorer so that things are
still executed efficiently eg. if a phrase or geo-distance query is wrapped
- filter out docs that don't have a high enough score using two-phase
iteration: this way the score is only checked when everything else matches
While doing these changes, I noticed that minScore was ignored when scores were
not needed and that explain did not take it into account, so I fixed these
issues as well.
This changes a couple of things:
Mappings are truly immutable. Before, each field mapper stored a
MappedFieldTypeReference that was shared across fields that have the same name
across types. This means that a mapping update could have the side-effect of
changing the field type in other types when updateAllTypes is true. This works
differently now: after a mapping update, a new copy of the mappings is created
in such a way that fields across different types have the same MappedFieldType.
See the new Mapper.updateFieldType API which replaces MappedFieldTypeReference.
DocumentMapper is now immutable and MapperService.merge has been refactored in
such a way that if an exception is thrown while eg. lookup structures are being
updated, then the whole mapping update will be aborted. As a consequence,
FieldTypeLookup's checkCompatibility has been folded into copyAndAddAll.
Synchronization was simplified: given that mappings are truly immutable, we
don't need the read/write lock so that no documents can be parsed while a
mapping update is being processed. Document parsing is not performed under a
lock anymore, and mapping merging uses a simple synchronized block.
This adds the required changes/checks so that the build can run on
FreeBSD.
There are a few things that differ between FreeBSD and Linux:
- CPU probes return -1 for CPU usage
- `hot_threads` cannot be supported on FreeBSD
From OpenJDK's `os_bsd.cpp`:
```c++
bool os::is_thread_cpu_time_supported() {
#ifdef __APPLE__
return true;
#else
return false;
#endif
}
```
So this API now returns (for each FreeBSD node):
```
curl -s localhost:9200/_nodes/hot_threads
::: {Devil Hunter Gabriel}{q8OJnKCcQS6EB9fygU4R4g}{127.0.0.1}{127.0.0.1:9300}
hot_threads is not supported on FreeBSD
```
- multicast fails in native `join` method - known bug:
https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=193246
Which causes:
```
1> Caused by: java.net.SocketException: Invalid argument
1> at java.net.PlainDatagramSocketImpl.join(Native Method)
1> at java.net.AbstractPlainDatagramSocketImpl.join(AbstractPlainDatagramSocketImpl.java:179)
1> at java.net.MulticastSocket.joinGroup(MulticastSocket.java:323)
1> at org.elasticsearch.plugin.discovery.multicast.MulticastChannel$Plain.buildMulticastSocket(MulticastChannel.java:309)
```
So these tests are skipped on FreeBSD.
Resolves#15562
In this commit we increase the queue size of the bulk pool in
BulkProcessorRetryIT to make it less sensitive.
As this test case should stress the pool so bulk processor needs to
back off but not so much that the backoff policy will give up at
some point (which is a valid condition), we still keep it below the
default queue size of 50.
Some tests, but in particular CodecTests, are slow because they test all
versions that ever existed even though they should only test supported
versions.
Today we throttle recoveries only for incoming recoveries. Nodes that have a lot
of primaries can get overloaded due to too many recoveries. To still keep that at bay
we limit the number of threads that are sending files to the target to overcome this problem.
The right solution here is to also throttle the outgoing recoveries that are today unbounded on
the master and don't start the recovery until we have enough resources on both source and target nodes.
The concurrency aspects of the recovery source also added a lot of complexity and additional threadpools
that are hard to configure. This commit removes the concurrent streamns notion completely and sends files
in the thread that drives the recovery simplifying the recovery code considerably.
Outgoing recoveries are not throttled on the master via a allocation decider.
We added this undocumented realtime setting as backup plan long ago
but to date we haven't had a situation where it was a problem. It's reducing
the number of filehandles in the NRT case dramatically and should always be enabled.
Today the logic to async - commit the translog is in every translog instance
itself. While the setting is a per index setting we manageing it per shard. This
polluts the translog code and can more easily be managed in IndexService.
Today we have two variants of translogs for indexing. We only recommend the buffered
one which also has a 20% advantage in indexing speed. This commit removes the option and defaults
to the buffered case. It also hard-wires the translog buffer to 8kb instead of 64kb. We used to
adjust that buffer based on if the shard is active or not, this code has also been removed and
instead we just keep an 8kb buffer arround.
This commit removes `index.translog.flush_threshold_ops` and `index.translog.disable_flush`
in favor of `index.translog.flush_threshold_size`. The number of operations is meaningless by itself and
can easily be turned into a size value with knowledge of the data. Disabling the flush is only useful in
tests and we can set the size value to a really high value. If users really need to do this they can
also apply a very high value like `1PB`.
DocumentMapperParser has both parse and parseCompressed methods. Except that the
parse methods are ONLY used from the unit tests. This commit removes the parse
method and moves all tests to parseCompressed so that they test more
realistically how mappings are managed.
Then I renamed parseCompressed to parse given that this is the only alternative
anyway.
Resolves conflicts between parent routing and alias routing with the following rule:
* The parent routing is ignored if there is an alias routing that matches the request.
Closes#3068
With this commit we implement a cancellation policy in
BulkProcessor which is aligned for the sync and the async case
and also document it.
Closes#14833.
It's important to close not matter what exception caused a tragic event. Today
we only check on IOException and AlreadyClosedExceptions. The test had a bug and
threw an IAE instead causing the translog not to be closed.
Nik Everett
Simon Willnauer
Britta Weber
Michael McCandless
Martijn van Groningen
Christoph Büscher
Jason Tedor
Ali Beyad
Michael McCandless
Adrien Grand
Simon Willnauer
Igor Motov
Boaz Leskes
Nicholas Knize
Yannick Welsch
Dave
Jim Ferenczi
Robert Muir
Daniel Mitterdorfer
Robert Muir
Lee Hinman
kwangsik
Jun Ohtani
Zachary Tong
Ryan Ernst
Areek Zillur
Colin Goodheart-Smithe