Today we keep track of how often filters are used at the index level in order
to decide whether they should be cached or not. This is an issue if you have
several shards of the same index on the same node as it will multiply statistics
by the number of shards that you have for this index on the node, which defeats
the purpose of waiting for a filter to be reused before caching them.
If the translog UUID is corrupted we should not convert it
to UTF-8 since it might be invalid. Instead we should compare
the UTF-8 byte representation directly.
This more consistent with the other logging it makes and since it can be used in many operations the output can be more verbose (without adding too much info as to who timed out exactly - which we can fix separately). If need be the caller of the observer can log a higher level message.
Closes#11722
In order to be more consistent with what they do, the query cache has been
renamed to request cache and the filter cache has been renamed to query
cache.
A known issue is that package/logger names do no longer match settings names,
please speak up if you think this is an issue.
Here are the settings for which I kept backward compatibility. Note that they
are a bit different from what was discussed on #11569 but putting `cache` before
the name of what is cached has the benefit of making these settings consistent
with the fielddata cache whose size is configured by
`indices.fielddata.cache.size`:
* index.cache.query.enable -> index.requests.cache.enable
* indices.cache.query.size -> indices.requests.cache.size
* indices.cache.filter.size -> indices.queries.cache.size
Close#11569
In order to be backwards compatible, indices created before 2.x must support
indexing of a unix timestamp and its configured date format. Indices created
with 2.x must configure the `epoch_millis` date formatter in order to
support this.
Relates #10971
Today we are very lenient in parsing the translog files. This is
actually not necessary since we have a clear run once upgrade path.
All files are converted into the new file name pattern such that we
only need to look at old file patterns in the context of the upgrade.
This commit makes parsing really strict with the exceptoin of the upgrade path.
This adds a new pipeline aggregation, the cumulative sum aggregation. This is a parent aggregation which must be specified as a sub-aggregation to a histogram or date_histogram aggregation. It will add a new aggregation to each bucket containing the sum of a specified metrics over this and all previous buckets.
Today we mark a translog as upgraded by adding a marker to the engine commit.
Yet, this commit was only added if there was no translog present before ie. only
if we have a fresh engine which is missing the entire point. Yet, this commit
adds a backwards index tests that ensures we can open old indices more than once
ie. mark the index as upgraded.
Closes#11858
In Lucene 5x the exception thrown when highlighter encounters a huge term
is a BytesRefHash.MaxBytesLengthExceededException but in Lucene 4x it is
wrapped in a RuntimeException. Therefore, it seems saver to unwrap this.
There was only a single actual "use" of close, for a threadlocal
in VersionFieldMapper. However, that threadlocal is completely
unnecessary, so this change removes the threadlocal and
close() altogether.
This commit consolidates several abstractions on the shard level in
ordinary classes not managed by the shard level guice injector.
Several classes have been collapsed into IndexShard and IndexShardGatewayService
was cleaned up to be more lightweight and self-contained. It has also been moved into
the index.shard package and it's operation is renamed from recovery from "gateway" to recovery
from "store" or "shard_store".
Closes#11847
This commit folds ShardRouting, ImmutableShardRouting and MutableShardRouting
into ShardRouting. All mutators are package private anyway today so it's just
unnecessary abstraction.
ShardRoutings are now frozen once they are added to the IndexRoutingTable
to prevent modifications outside of the allocation code.
This commit makes the get and search APIs always return `_parent`, `_routing`,
`_timestamp` and `_ttl` in addition to `_id` and `_type`. This way, consumers
always have all required information in order to reindex a document.
Currently the SnapshotsService is concerned with both maintaining the global snapshot lifecycle on the master node as well as responsible for keeping track of individual shards on the data nodes. This refactoring separates two areas of concerns by moving all shard-level operations into a separate SnapshotShardsService.
Closes#11756
Currently the filter cache is configured to have a maximum size in bytes of 10%
of the JVM memory, and a maximum number of cached filters (across all segments
of all shard on the same node) of 100000. I would like to change the latter to
a more reasonable value of 1000.
Given that we track the most 256 most recently used filters per index and only
cache those that have been seen 5 times or more, a single index cannot have more
than 50 hot filters, so a maximum number of cached filters of 1000 per node
should be more than necessary.
Today, we have scheduled reroute that kicks every 10 seconds and checks if a
reroute is needed. We use it when adding nodes, since we don't reroute right
away once its added, and give it a time window to add additional nodes.
We do have recover after nodes setting and such in order to wait for enough
nodes to be added, and also, it really depends at what part of the 10s window
you end up, sometimes, it might not be effective at all. In general, its historic
from the times before we had recover after nodes and such.
This change removes the 10s scheduling, simplifies RoutingService, and adds
explicit reroute when a node is added to the system. It also adds unit tests
to RoutingService.
closes#11776
Since elasticsearch doesn't shade artifacts anymore (see #11522), the dependencies list for RPM/DEB must be updated. Now we package all maven libs by default except the generated -shaded/-tests/-test-cours JARs and slf4j-api (marked as optionnal).
We currently are very lax about allowing data types to conflict for the
same field name, across document types. This change makes the underlying
map in MapperService a 1-1 map of field name to field type, and throws
exception when new types are not compatible.
To still allow changing a type, with parameters that are allowed to be
changed, but for a field that exists in multiple types, a new parameter
to index creation and put mapping API is added: update_all_types.
This defaults to false, and the exception messages suggest using
this parameter when trying to modify a setting that is allowed to be
modified but is being limited by this restriction.
There are also a couple changes which try to base fields from new types
for dynamic mappings, and root mappers, on existing settings. For
dynamic mappings this is important if the dynamic defaults have been
changed. For root mappings, this is mostly just for backcompat when
pre 2.0 root mappers could have their field type changed.
fixes#8871
we currently don't expose this.
This adds the following to the OS section of `_nodes`:
```
"os": {
"name": "Mac OS X",
...
}
```
and the following to the OS section of `_cluster/stats`:
```
"os": {
...
"names": [
{
"name": "Mac OS X",
"count": 1
}
],
...
},
```
Closes#11807
This is a follow up to #8143 and #6730 for _timestamp. It removes
support for `path`, as well as any field type settings, and
enables docvalues for _timestamp, for 2.0. Users who need to
adjust these settings can use a date field.
- Fixes tests, and removes a few special snowflake, fragile tests.
- Removes concrete implementation of predict() and moves it into
each model so that the logic is clearer. Because there is some
shared checks/assertions, those remain in predict() and the main
prediction happens in doPredict()
The commit about adding cluster health response features also removed
accidentally some functionality, that resulted in wrong instanceof checks
in InternalClusterService and thus in test failures because the cluster
state task that was added via an anonymous was missing the cast.
This commit readds the abstract class with slight renaming.
Commit id was: 88f8d58c8b
If we mark the shard as being in POST_RECOVERY before the percolator
is fully set up we might expose it to the user as fully searchable before
all queries are loaded. This can lead to wrong results especially in tests
when a shard is concurrently marked as STARTED.
This commit also removes unneded abstractions on IndexShard where readoperations
should be allowed when the purose is a write.
In order to get a quick overview using by simply checking the cluster state
and its corresponding cat API, the following two attributes have been added
to the cluster health response:
* task max waiting time, the time value of the first task of the
queue and how long it has been waiting
* active shards percent: The percentage of the number of shards that are in
initializing state
This makes the cluster health API handy to check, when a fully restarted
cluster is back up and running.
Closes#10805
The change makes rest-spec-api a project in the same way as we build dev-tools. it packages the tests and api in a bundle using the maven-remote-resources-plugin and uses the same plugin in the plugins and core pom to unpack the rest-api-spec into the target directory and references the rest tests there in the test resources.
The main stimulus for this change is that for those using Eclipse the current build does not work. After running `mvn eclipse:eclipse` the Eclipse IDE errors because the rest-api-spec is outside of the project scope, meaning that every time the command is run (required whenever any dependencies change), the class path of all the projects has to be manually fixed.
This fixes an issue to allow for negative unix timestamps.
An own printer for epochs instead of just having a parser has been added.
Added docs that only 10/13 length unix timestamps are supported
Added docs in upgrade documentation
Fixes#11478
Tests relying on sleeps and latch timeouts are prone to weird timing issues
and hard to read / understand error messages. This commit moves towards a more
deterministic error model and replaces empty fails with real exceptions.
Some repository verification exceptions are currently only returned to the users but not logged on the nodes where the exceptions occurred, which makes troubleshooting difficult.
Closes#11760
For #8871, we need to be able to check field types are compatible,
without comparing FieldMappers. This change moves the simulation
checks (which generate merge conflicts) for any properties of
MappedFieldTypes into a new method, validateCompatible.
This also simplifies the merge code which merges settings
between the old and new fieldtypes. Previously, each subclass
of FieldMapper would have to set its own fieldtype settings.
However, now that we have .clone(), which perfectly copies
all properties (with subclasses accounted for), we can now
do a simple clone when merging.
Finally, this fixes a subtle bug in merging, in which if
merging has conflicts, and we were not simulating, we would
still update the field type, even though it was not compatible!
NOTE: there is one test failure I am trying to track down with
timestamp merging. Otherwise, all tests pass.
Currently, we delete the shard _state file on engine failure.
This behaviour does not persist the engine failure reason for later inspection.
This commit marks the shard store as corrupted instead of deleting
the _state file to ensure the store index can not be opened after and
the engine failure is persisted.
CommonTermsQueryParser does not check for disable_coords, only for
disable_coord. Yet the builder only outputs disable_coords, leading to
disabling the coordination factor to be ignored in the Java API.
Closes#11730Closes#11780
When using compression over the network, you might sometimes see warnings that
the stream was not fully read. This is because DeflaterOutputStream adds an
end-of-stream marker. When deserializing, we need to poll for one byte using
InputStream.read() to make sure to decode this EOS marker.
For the record, it does not strike all the time today because we perform
buffering when decompressing to avoid performing too many JNI calls, but it
is easy to make this warning happen all the time by decreasing the size of
the buffer we use.
Close#11748
Need to reset the registered setting in order to make sure the nex round will capture the right delay interval
also randomize setting and name the setting properly
closes#11759
Added several classes to support expressions being used for numerical
calculations in aggregations. Expressions will still not compile
when used with mapping and update script contexts.
Closes#11596Closes#11689
Allow to set delayed allocation timeout on unassigned shards when a node leaves the cluster. This allows to wait for the node to come back for a specific period in order to try and assign the shards back to it to reduce shards movements and unnecessary relocations.
The setting is an index level setting under `index.unassigned.node_left.delayed_timeout` and defaults to 0 (== no delayed allocation). We might want to change the default, but lets do it in a different change to come up with the best value for it. The setting can be updated dynamically.
When shards are delayed, a log message with "info" level will notify how many shards are being delayed.
An implementation note, we really only need to care about delaying allocation on unassigned replica shards. If the primary shard is unassigned, anyhow we are going to wait for a copy of it, so really the only case to delay allocation is for replicas.
close#11712
Colin Goodheart-Smithe
Adrien Grand
Simon Willnauer
Martijn van Groningen
Masaru Hasegawa
Boaz Leskes
Robert Muir
Alexander Reelsen
Britta Weber
Ryan Ernst
Robert Muir
Tanguy Leroux
Jörg Prante
David Pursehouse
Igor Motov
Wolfgang Karall
Shay Banon
Clinton Gormley
javanna
Zachary Tong
Simon Willnauer
Johannes Barthelmes
Areek Zillur
Alex Ksikes
Iarwa1n
Jack Conradson