At the final stage of a relocation, during the final flip of the states, a search request might hit a node that would then execute it on a shard that has already relocated.
For this, we need to execute broadcast and search operations against initializing shards as well, but only as a last resort. The operation will be rejected if not applicable (i.e. IndexShard#searcher() checked for read allowed).
Note, this requires careful though about which failures we send back. If we try and initializing shard and it fails, its failure should not override an actual failure of an active shard.
Also, removed an atomic integer used in broadcast request and use a similar shard index trick we now have in our search execution.
closes#3427
UpdateNumberOfReplicasTests#simpleUpdateNumberOfReplicasTests is very
fragile due to executing searches based on dated knowledge of
the cluster state and calling shards that have been relocating away in
the mean time. A fix is on the way.
* Changed the response to include the alias as part of each match.
* Added `percolate_format=ids` query string option to just serialize the ids in the rest response.
* Added support for multiple indices in the percolate api.
Closes#3420
This commit introduces near realtime suggestions. For more information about
its usage refer to github issue #3376
From the implementation point of view, a custom AnalyzingSuggester is used
in combination with a custom postingsformat (which is not exposed to the user
anywhere for him to use).
Closes#3376
We have an optimization where we try to delay reroute after we processed the shard started events to try and combine a few into the same event. With teh queueing of shard started events in places, we don't need to do it, and we can reroute right away, which will actually reduce the amount of cluster state events we send.
This will also have a nice side effect of not missing on "waitForRelocatingShards(0)" on cluster health checks since relocations will happen right away.
closes#3417
- checking routing table taken from same (up-to-date) cluster state
- added @Slow annotation
- forced cluster reroute when needed
- changed order of assertions so that if it fails again it's easier to understand why
this will not happen now, but in the future, if data nodes will only be in memory (including translog and such), then we need to fire the deleted events
StringReader is synchronized although input streams should always be consumed
by a single thread at a time. FastStringReader on the other hand is completely
thread unsafe.
Closes#3411
Lucene 4.4 introduced the tokenStream(String fieldName, String text) helper
method which nicely reuses the string reader. Whenever analyzing Strings, we
should use this method instead of using the Reader method with a
(Fast)StringReader.
Closes#3409
The following changes improved alias creation:
* Moved away from ImmutableMap to JCF's UnmodifiableMap. The ImmutableMap always made a copy, whereas the UnmodifiableMap just wraps the target map.
* Reducing the number of maps being created during the creation of MetaData and IndexMetadata.
* Changed IndexAliasesService's aliases from a copy on write ImmutableMap to ConcurrentMap.
Closes#3410
If it's acknowledged all the other nodes already hold the latest cluster state, thus we can search on the closed index and be sure we get an error back
The open/close index api now waits for an acknowledgement from all the other nodes before returning its response, till the timeout (configurable, default 10 secs) expires. The returned acknowledged flag reflects whether the cluster state change was acknowledged by all the nodes or the timeout expired before.
Closes#3400
- remove Digest/Hex, only used in test, directly use MD5 there
- don't use Closeables and use IOUtils
- remove duplicate Classes usage and unused LongsLAB
Shay Banon
Luca Cavanna
Simon Willnauer
Martijn van Groningen
Alexander Reelsen
Adrien Grand
Peter Moberg
Boaz Leskes