Currently, we do not allow reads on shards which are in POST_RECOVERY which
unfortunately can cause search failures on shards which just recovered if there no replicas (#9421).
The reason why we did not allow reads on shards that are in POST_RECOVERY is
that after relocating a shard might miss a refresh if the node that executed the
refresh is behind with cluster state processing. If that happens, a user might execute
index/refresh/search but still not find the document that was indexed.
We changed how refresh works now in #13068 to make sure that shards cannot miss a refresh this
way by sending refresh requests the same way that we send write requests.
This commit changes IndexShard to allow reads on POST_RECOVERY now.
In addition it adds two test:
- test for issue #9421 (After relocation shards might temporarily not be searchable if still in POST_RECOVERY)
- test for visibility issue with relocation and refresh if reads allowed when shard is in POST_RECOVERY
closes#9421
DateHistogramIT has suite scope and therefore must take care of cleaning up after
each test. Otherwise we cannot run tests in several times with tests.iters.
The implementation this commit replaces was almost k-NN regression with
k=2, but had two bugs: (a) it depends on the empirical raw estimates
being in strictly non-decreasing order for the binary search (which they
are not); and (b) it weights the biases positively with increased
distance from the corresponding raw estimate.
“HyperLogLog in Practice” leaves the choice of exact algorithm here
fairly vague, just noting: “We use k-nearest neighbor interpolation to
get the bias for a given raw estimate (for k = 6).” The majority of
other open source HyperLogLog++ implementations appear to use k-NN
regression with uniform weights (and generally k = 6). Uniform
weighting does decrease variance, but also introduces bias at the domain
extrema. This problem, plus the use of the word “interpolation” in the
original paper, suggests (inverse) distance-weighted k-NN, as
implemented here.
Since #13068 refresh and flush requests go to the primary first and are then replicated.
One difference to before is though that if a shard is not available (INITIALIZING for example)
we wait a little for an indexing request but for refresh we don't and just give up immediately.
Before, refresh requests were just send to the shards regardless of what their state is.
In tests we sometimes create an index, issue an indexing request, refresh and
then get the document. But we do not wait until all nodes know that all primaries have ben assigned.
Now potentially one node can be one cluster state behind and not know yet that
the shards have ben started. If the refresh is executed through this node then the
refresh request will silently fail on shards that are started already because from
the nodes perspective they are still initializing. As a consequence, documents
that expected to be available in the test are now not.
Example test failures are here: http://build-us-00.elastic.co/job/elasticsearch-20-oracle-jdk7/395/
This commit changes the timeout to 1m (default) to make sure we don't miss shards
when we refresh. This will trigger the same retry mechanism as for indexing requests.
We still have to make a decision if this change of behavior is acceptable.
see #13238
From a user perspective, the main benefit from this upgrade is that the new
Lucene53Codec has disk-based norms. The elasticsearch directory has been fixed
to load these norms through mmap instead of nio.
Other changes include the removal of `max_thread_states`, the fact that
PhraseQuery and BooleanQuery are now immutable, and that deleted docs are now
applied on top of the Scorer API.
This change introduces a couple of `AwaitsFix`s but I don't think it should
hold us from merging.
This allows reducing privileges with doPrivileged to work,
otherwise it will fail with NPE.
In general, if some code wants to do that, let it. The null
check is needed, even though ProtectionDomain(CodeSource, PermissionCollection)
is more than a bit misleading: "the current Policy will not be consulted".
Additionally add a defensive check for location, since the docs
there are even more confusing: https://bugs.openjdk.java.net/browse/JDK-8129972
The jdk policy impl has both these checks.
Virtualbox is the default virtualization provier for vagrant but folks
override that from time to time. If they do then the build will fail because
the boxes used by the build don't usually support non-virtualbox providers.
Closes#13217
Prior to 2.0 we summed up the available space on all disk on a node
due to the raid-0 like behavior. Now we don't do this anymore and use the
min & max disk space to make decisions.
Closes#13106
Isabel Drost-Fromm
Britta Weber
Robert Muir
Robert Muir
Ryan Ernst
Simon Willnauer
Nik Everett
Adrien Grand
Marshall Bockrath-Vandegrift
Clinton Gormley
Jason Tedor
Isabel Drost-Fromm
xuzha
Martijn van Groningen
Tanguy Leroux