A stuck peer recovery in #40913 reveals that we indefinitely retry on
new cluster states if indexing translog operations hits a mapper
exception. We should not wait and retry if the mapping on the target is
as recent as the mapping that the primary used to index the replaying
operations.
Relates #40913
This commit refactors GeoHashUtils class into a new Geohash utility class located in the ES geo library. The intent is to not only better control what geo methods are whitelisted for painless scripting but to clean up the geo utility API in general.
Motivated by slow snapshot deletes reported in e.g. #39656 and the fact that these likely are a contributing factor to repositories accumulating stale files over time when deletes fail to finish in time and are interrupted before they can complete.
* Makes snapshot deletion async and parallelizes some steps of the delete process that can be safely run concurrently via the snapshot thread poll
* I did not take the biggest potential speedup step here and parallelize the shard file deletion because that's probably better handled by moving to bulk deletes where possible (and can still be parallelized via the snapshot pool where it isn't). Also, I wanted to keep the size of the PR manageable.
* See https://github.com/elastic/elasticsearch/pull/39656#issuecomment-470492106
* Also, as a side effect this gives the `SnapshotResiliencyTests` a little more coverage for master failover scenarios (since parallel access to a blob store repository during deletes is now possible since a delete isn't a single task anymore).
* By adding a `ThreadPool` reference to the repository this also lays the groundwork to parallelizing shard snapshot uploads to improve the situation reported in #39657
* Thanks to #39793 dynamic mapping updates don't contain blocking operations anymore so we don't have to manually put the mapping in this test and can keep it a little simpler
* Add Restore Operation to SnapshotResiliencyTests
* Expand the successful snapshot test case to also include restoring the snapshop
* Add indexing of documents as well to be able to meaningfully verify the restore
* This is part of the larger effort to test eventually consistent blob stores in #39504
* The check doesn't add much if anything practically, since the S3 repository is eventually consistent and we only log the non-existence of a blob anyway
* We don't do the check on writes for this very reason and documented it as such
* Removing the check saves one API call per single delete speeding up the deletion process and lowering costs
Today the `_field_caps` API returns the list of indices where a field
is present only if this field has different types within the requested indices.
However if the request is an index pattern (or an alias, or both...) there
is no way to infer the indices if the response contains only fields that have
the same type in all indices. This commit changes the response to always return
the list of indices in the response. It also adds a way to retrieve unmapped field
in a specific section per field called `unmapped`. This section is created for each field
that is present in some indices but not all if the parameter `include_unmapped` is set to
true in the request (defaults to false).
* Introduce Delegating ActionListener Wrappers
* Dry up use cases of ActionListener that simply pass through the response or exception to another listener
* Name Snapshot Data Blobs by UUID
* There is no functional reason why we need incremental naming for these files but
* As explained in #38941 it is a possible source of corrupting the repository
* It wastes API calls for the list operation
* Is just needless complication
* Since we store the exact names of the data blobs in all the metadata anyway, we can make this change without any BwC considerations
* Even on the worst case scenario of a downgrade the functionality would continue working since the incremental names wouldn't conflict with the uuids and the number parsing for finding the next incremental name suppresses the exception when encountring a non-numeric value after the double underscore prefix
In order to support empty action metadata in the first msearch item,
we need to remove support for prepending msearch request body with an
empty line, which prevents us from parsing the empty line as action
metadata for the first search item.
Relates to #41011
The test is testing the java time API and fails in case it hits daylight saving time changes.
Java time has the right implementation and we don't need to test this.
more details on how the test was affected by the DST change on this comment
closes#39617
backport(#41493)
* Due to #40866 one of the two parallel bulk requests can randomly be
rejected outright when the write queue is full already, we can catch
this situation and ignore it since we can still have the rejection for
the dynamic mapping udate for the other reuqest and it's somewhat rare
to run into this anyway
* Closes#41363
Adds some validation to prevent duplicate source names from being
used in the composite agg.
Also refactored to use a ConstructingObjectParser and removed the
private ctor and setter for sources, making it mandatory.
* The problem here is that if we run into a corrupted index-N file, instead of generating a new index-(N+1) file, we instead set the newest index generation to -1 and thus tried to create `index-0`
* If `index-0` is corrupt, this prevents us from ever creating a new snapshot using the broken shard, because we are unable to create `index-0` since it already exists
* Fixed by still using the index generation for naming the next index file, even if it was a broken index file
* Added test that makes sure restoring as well as snapshotting on top of the broken shard index file work as expected
* closes#41304
* The test fails for the retry backoff enabled case because the retry handler in the bulk processor hasn't been adjusted to account for #40866 which now might lead to an outright rejection of the request instead of its items individually
* Fixed by adding retry functionality to the top level request as well
* Also fixed the duplicate test for the HLRC that wasn't handling the non-backoff case yet the same way the non-client IT did
* closes#41324
Today we assert that there are no operations in flight in this test. However we
will sometimes be in a situation where the operations are blocked, and we
distinguish these cases since #41271 causing the assertion to fail. This commit
addresses this by allowing operations to be blocked sometimes after a primary
promotion.
Fixes#41333.
The `composite` aggregation maps unknown fields as numerics, this means that
any `after` value that is set on a query with an unmapped field on some indices
will fail if the provided value is not numeric. This commit changes the default
value source to use keyword instead in order to be able to parse any type of after
values.
The `_id` field uses a binary encoding to index terms that is not compatible with
the utf8 automaton that the unified highlighter creates to reanalyze the input.
For these reason this commit ignores terms that target the `_id` field when
`require_field_match` is set to false.
Closes#37525
With the removal of the `_all` field the `mlt` query cannot infer a field name
to use to analyze the provided (un)like text if the `fields` parameter is not
explicitly set in the query and the `index.query.default_field` is not changed
in the index settings (by default it is set to `*`). For this reason the like text
is ignored and queries are only built from the provided document ids.
This change fixes this bug by throwing an error if the fields option is not set
and the `index.query.default_field` is equals to `*`. The error is thrown only
if like or unlike texts are provided in the query.
This fixes an issue where every N seconds a slow search request is triggered
since the searcher access time is not set unless the shard is idle. This change
moves to a more pro-active approach setting the searcher as accessed all the time.
* Bulk requests can be thousands of items large and take more than O(10ms) time to handle => we should not handle them on the transport threadpool to not block select loops
* relates #39128
* relates #39658
Today we do not distinguish "no operations in flight" from "operations are
blocked", since both return `0` from `IndexShard#getActiveOperationsCount()`.
We therefore cannot assert that every `TransportReplicationAction` performs its
actions under permit(s). This commit fixes this by returning
`IndexShard#OPERATIONS_BLOCKED` if operations are blocked, allowing these two
cases to be distinguished.
The date_histogram internally converts obsolete timezones (such as
"Canada/Mountain") into their modern equivalent ("America/Edmonton").
But rollup just stored the TZ as provided by the user.
When checking the TZ for query validation we used a string comparison,
which would fail due to the date_histo's upgrading behavior.
Instead, we should convert both to a TimeZone object and check if their
rules are compatible.
The `ignore_malformed` option currently works on numeric fields only when the
bad value isn't a string value but not if it is a boolean. In this case we get a
parsing error from the xContent parser which we need to catch in addition to the
field mapper.
Closes#11498
Today the `?preference=custom_string_value` search preference will only change
its choice of a shard copy if something changes the `IndexShardRoutingTable`
for that specific shard. Users can use this behaviour to route searches to a
consistent set of shard copies, which means they can reliably hit copies with
hot caches, and use the other copies only for redundancy in case of failure.
However we do not assert this property anywhere, so we might break it in
future.
This commit adds a test that shows that searches are routed consistently even
if other indices are created/rebalanced/deleted.
Relates https://discuss.elastic.co/t/176598, #41115, #26791
Today we always trim unsafe commits (whose max_seq_no >= global
checkpoint) before starting a read-write or read-only engine. This is
mandatory for read-write engines because they must start with the safe
commit. This is also fine for read-only engines since most of the cases
we should have exactly one commit after closing an index (trimming is a
noop). However, this is dangerous for following indices which might have
more than one commits when they are being closed.
With this change, we move the trimming logic to the ctor of InternalEngine
so we won't trim anything if we are going to open a read-only engine.
Currently enabling profiling disables top-hits optimizations, which is
unfortunate: it would be nice to be able to notice the difference in method
counts and timings depending on whether total hit counts are requested.
`Node#close` is pretty hard to rely on today:
- it might swallow exceptions
- it waits for 10 seconds for threads to terminate but doesn't signal anything
if threads are still not terminated after 10 seconds
This commit makes `IOException`s propagated and splits `Node#close` into
`Node#close` and `Node#awaitClose` so that the decision what to do if a node
takes too long to close can be done on top of `Node#close`.
It also adds synchronization to lifecycle transitions to make them atomic. I
don't think it is a source of problems today, but it makes things easier to
reason about.
Today we check if an index has broken settings when checking if an index
needs to be upgraded. However, it can be the case that an index setting
became broken even if an index is already upgraded to the current
version if the user removed a plugin (or downgraded from the default
distribution to the non-default distribution) while on the same version
of Elasticsearch. In this case, some registered settings would go
missing and the index would now be broken. Yet, we miss this check and
instead of archiving the settings, the index becomes unassigned due to
the missing settings. This commit addresses this by checking for broken
settings whether or not the index is upgraded.
One of the two #getCorrections methods is only used in tests, so we can move
it and any of the required helper methods to that test. Also reducing the
visibility of several methods to package private since the class isn't used
elsewhere outside the package.
Today we erroneously look for a node setting called `readonly` when deciding
whether or not to create a missing directory in a filesystem repository. This
change fixes this by using the repository setting instead.
Closes#41009
Relates #26909
In `TransportRolloverAction` before doing rollover we resolve
source index name (write index) from the alias in the rollover request.
Before evaluating the conditions and executing rollover action, we
retrieve stats, but to do so we used the source index name
resolved from the alias instead of alias from the index.
This fails when the user is assigned a role with index privilege on the
alias instead of the concrete index. This commit fixes this by using
the alias from the request.
After this change, verified that when we retrieve all the stats (including write + read indexes)
we are considering only source index.
Closes#40771
The unified highlighter returns the first sentence of the text when number_of_fragments
is set to 0 (full highlighting). This is a legacy of the removed postings highlighter
that was based on sentence break only. This commit changes this behavior in order
to respect the provided no_match_size value when number_of_fragments is set to 0.
This means that the behavior will be consistent for any value of the number_of_fragments option.
Closes#41066
* Adds Bulk delete API to blob container
* Implement bulk delete API for S3
* Adjust S3Fixture to accept both path styles for bulk deletes since the S3 SDK uses both during our ITs
* Closes#40250
Today the blended term query detects if a term exists in a field by looking at the term statistics in the index.
However the value to indicate that a term has no occurence in a field have changed in Lucene. A non-existing term now returns
a doc and total term frequency of 0. Because of this disrepancy the blended term query picks 0 as the minimum frequency for a term
even if other fields have documents for this terms. This confuses the term queries that the blending creates since some of them
contain a custom state that indicates a frequency of 0 even though the term has some occurence in the field. For these terms an exception
is thrown because the term query always checks that the term state's frequency is greater than 0 if there are documents associate to it.
This change fixes this bug by ignoring terms with a doc freq of 0 when the blended term query picks the minimum term frequency among the
requested fields.
Closes#41118
`TransportReplicationAction.AsyncPrimaryAction#createReplicatedOperation`
exists so it can be overridden in tests. This commit re-works these tests to
use a real `ReplicationOperation` and inlines the now-unnecessary method.
Relates #40706.
Today we fail to join a Zen2 cluster if the cluster UUID does not match our
own, but we do not perform the same validation when joining a Zen1 cluster.
This means that a Zen2 node will pass join validation and be added to a Zen1
cluster but will reject all cluster states from the master.
Relates #37775
Currently we throw an error when a range querys minimum value exceeds the
maximum value due to the fact that they are neighbouring values and both upper
and lower value are excluded from the interval.
Since this is a condition that the user usually doesn't specify conciously (at
least in the case of float and double values its difficult to see which values
are adjacent) we should ignore those "wrong" intervals and create a
MatchNoDocsQuery in those cases.
We should still throw errors with an actionable message if the user specifies
the query interval in a way that min value > max value. This PR adds those
checks and tests for those cases.
Closes#40937
This is related to #36652. We intend to deprecate a number of transport
settings in 7.x and remove them in 8.0. This commit removes the string
usages of these settings.
Pipelines require single-valued agg or a numeric to be returned.
If they don't get that, they throw an exception. Unfortunately, this
exception text is very confusing to users because it usually arises
from pathing "through" multiple terms aggs. The final target is a numeric,
but it's the intermediary aggs that cause the problem.
This commit adds the current agg name to the exception message
so the user knows which "level" is the issue.
Full text queries ignore unmapped fields since https://github.com/elastic/elasticsearch/issues/41022
even if all fields in the query are unmapped.
This change makes sure that we ignore unmapped fields only if they are mixed
with mapped fields and returns a MatchNoDocsQuery otherwise.
Closes#41022
This change adds either ToXContentObject or ToXContentFragment to classes
directly implementing ToXContent currently. This helps in reasoning about
whether those implementations output full xcontent object or just fragments.
Relates to #16347
When a polygon contains a self-intersection due to have twice the same point in no-consecutive position, the polygon builder tries to split the polygon. During the split one of the polygons become invalid as it is not closed and an error is thrown which is not related to the real issue.
We detect this situation now and throw a more meaningful error.
Today a new replica of a closed index does not have a safe commit
invariant when its engine is opened because we won't initialize the
global checkpoint on a recovering replica until the finalize step. With
this change, we can achieve that property by creating a new translog
with the global checkpoint from the primary at the end of phase 1.
This helps avoid memory issues when computing deep sub-aggregations. Because it
should be rare to use sub-aggregations with significant terms, we opted to always
choose breadth first as opposed to exposing a `collect_mode` option.
Closes#28652.
Since #40249, we always reinitialize max_seq_no_of_updates to max_seq_no
when a promoting primary restores history regardless of whether it did
rollback previously or not.
Closes#40929
This commit removes the settings member variable from Node.
This member made it confusing which settings should actually be looked
at. Now all settings are accessed through the final environment.
A small refactoring that removes the primaryTerm field from ReplicasProxy and
instead passes it directly in to the methods that need it. Relates #40706.
This is a dependency of #39504
Motivation:
By refactoring `TransportShardBulkAction#shardOperationOnPrimary` to async, we enable using `DeterministicTaskQueue` based tests to run indexing operations. This was previously impossible since we were blocking on the `write` thread until the `update` thread finished the mapping update.
With this change, the mapping update will trigger a new task in the `write` queue instead.
This change significantly enhances the amount of coverage we get from `SnapshotResiliencyTests` (and other potential future tests) when it comes to tracking down concurrency issues with distributed state machines.
The logical change is effectively all in `TransportShardBulkAction`, the rest of the changes is then simply mechanically moving the caller code and tests to being async and passing the `ActionListener` down.
Since the move to async would've added more parameters to the `private static` steps in this logic, I decided to inline and dry up (between delete and update) the logic as much as I could instead of passing the listener + wait-consumer down through all of them.
* Replace usages RandomizedTestingTask with built-in Gradle Test (#40978)
This commit replaces the existing RandomizedTestingTask and supporting code with Gradle's built-in JUnit support via the Test task type. Additionally, the previous workaround to disable all tasks named "test" and create new unit testing tasks named "unitTest" has been removed such that the "test" task now runs unit tests as per the normal Gradle Java plugin conventions.
(cherry picked from commit 323f312bbc829a63056a79ebe45adced5099f6e6)
* Fix forking JVM runner
* Don't bump shadow plugin version
It could be that we try to shutdown the executor pool before all the
listeners have been invoked. It can happen that one was not invoked if
it timed out and was in the process of being notified that it timed out
on the executor. If we do this shutdown then, a listener will be met
with rejected execution exception. To address this, we first wait until
all listeners have been notified (or timed out) before proceeding with
shutting down the executor.
Relates #40970
Added documentation for node repurpose tool and included documentation on how to repurpose nodes safely. Adjusted order of tools in `elasticsearch-node` tool since the repurpose tool is most likely to be used.
Co-Authored-By: David Turner <david.turner@elastic.co>
Today if `cluster.routing.rebalance.enable: none` then rebalancing is disabled,
but we still execute `balanceByWeights()` and perform some rather expensive
calculations before discovering that we cannot rebalance any shards. In a large
cluster this can make cluster state updates occur rather slowly. With this
change we check earlier whether rebalancing is globally disabled and, if so,
avoid the rebalancing process entirely.
Relates #40942 which was reverted because of egregiously faulty tests.
The number of user data attributes of an index commit has increased
from 6 to 8, but we forgot to adjust. This change increases the initial
size of that map to avoid resizing.
Reducing some methods scope and marking them as static where possible. Removing
"alias" support from AnalysisRegistry#produceAnalyze and changing that method to
return a NamedAnalyzer instead of having a side effect on the analyzer map passed in.
Also, CustomAnalyzerProvider doesn't seem to need the `environment` field.
Today if `cluster.routing.rebalance.enable: none` then rebalancing is disabled,
but we still execute `balanceByWeights()` and perform some rather expensive
calculations before discovering that we cannot rebalance any shards. In a large
cluster this can make cluster state updates occur rather slowly. With this
change we check earlier whether rebalancing is globally disabled and, if so,
avoid the rebalancing process entirely.
The phrase suggesters have an option to remove terms that have
a frequency lower than a provided min_doc_freq. However this value is
overwritten by the frequency of the original term in the popular mode.
This change ensures that we keep the maximum value between the provided
min_doc_value and the original term frequency as a threshold to select
candidates.
Fixes#16764
Today we are strict when parsing build flavor and types off the
wire. This means that if a later version introduces a new build flavor
or type, an older version would not be able to parse what that new
version is sending. For a practical example of this, we recently added
the build type "docker", and this means that in a rolling upgrade
scenario older nodes would not be able to understand the build type that
the newer node is sending. This breaks clusters and is bad. We do not
normally think of adding a new enumeration value as being a
serialization breaking change, it is just not a lesson that we have
learned before. We should be lenient here though, so that we can add
future changes without running the risk of breaking ourselves
horribly. It is either that, or we have super-strict testing
infrastructure here yet still I fear the possibility of mistakes. This
commit changes the parsing of build flavor and build type so that we are
still strict at startup, yet we are lenient with values coming across
the wire. This will help avoid us breaking rolling upgrades, or clients
that are on an older version.
If the transport service is stopped, likely because we are shutting
down, and a retention lease background sync fires the logs will display
a warn message and stacktrace. Yet, this situaton is harmless and can
happen as a normal course of business when shutting down. This commit
suppresses the log messages in this case.
The `getIndexShard()` and `sendReplicaRequest()` methods in
TransportReplicationAction are effectively only used to customise some
behaviour in tests. However there are other ways to do this that do not cause
such an obstacle to separating the TransportReplicationAction into its two
halves (see #40706).
This commit removes these customisation points and injects the test-only
behaviour using other techniques.
Many gradle projects specifically use the -try exclude flag, because
there are many cases where auto-closeable resource ignore is never
referenced in body of corresponding try statement. Suppressing this
warning specifically in each case that it happens using
`@SuppressWarnings("try")` would be very verbose.
This change removes `-try` from any gradle project and adds it to the
build plugin. Also this change removes exclude flags from gradle projects
that is already specified in build plugin (for example -deprecation).
Relates to #40366
Primary-replica resync in a mixed-cluster between 6.x and 5.6 can send
operations without sequence number to a replica which already processed
operations with sequence number. This leads to the failure of that
replica for we trip the sequence number assertion when writing resync
operations without sequence number to translog.
This change rejects an illegal combination of flush parameters where
force is true, but wait_if_ongoing is false. This combination is trappy
and should be forbidden.
Closes#36342
If there's an ongoing flush triggered by the translog flush threshold,
we may fail to execute a flush because waitIfOngoing is false by
default.
Relates to #36342
If a refresh, which is scheduled by the setting change, executes after
the index-2 operation and win the refresh race (i.e., maybeRefresh) with
the scheduledRefresh that we are going to check, then the latter will
return false.
Closes#39565
Relates #39462
PR #40387
A user reported that the same query that takes ~900ms when querying an index
pattern only takes ~50ms when only querying indices that have matches. The
query is a date range query and we confirmed that the `can_match` phase works
as expected. I was able to reproduce this issue locally with a single node: with
900 1-shard indices, a query to an index pattern that matches all indices runs
in ~90ms while a query to the only index that has matches runs in 0-1ms.
This ended up not being related to the `can_match` phase but to the cost of
resolving aliases when querying an index pattern that matches lots of indices.
In that case, we first resolve the index pattern to a list of concrete indices
and then for each concrete index, we check whether it was matched through an
alias, meaning we might have to apply alias filters. Unfortunately this second
per-index operation runs in linear time with the number of matched concrete
indices, which means that alias resolution runs in O(num_indices^2) overall.
So queries get exponentially slower as an index pattern matches more indices.
I reorganized alias resolution into a one-step operation that runs in linear
time with the number of matches indices, and then a per-index operation that
runs in linear time with the number of aliases of this index. This makes alias
resolution run is O(num_indices * num_aliases_per_index) overall instead. When
testing the scenario described above, the `took` went down from ~90ms to ~10ms.
It is still more than the 0-1ms latency that one gets when only querying the
single index that has data, but still much better than what we had before.
Closes#40248
If a field `field_name` was missing in a document,
doc['field_name'].get(0) incorrectly retrieved
a value of the previously accessed document.
This happened because `get(int index)` function
was just accessing `values[index]` without
checking the number of values - `count`.
This PR fixes this.
There were some test failures caused by the background retention lease sync running on a relocated
primary. This commit fixes the situation that triggered the assertion and reactivates the failing test.
Closes#40731
The Eclipse compiler (4.10, Photon) cannot build this test because it cannot
correctly infer the type arguments of the functions. Explicitely adding them
helps in this case.
This change adds the following internal refactorings:
* wraps input analyzers into an unmodifiable map in IndexAnalyzers ctor
* removes duplicated indexSetting in IndexAnalyzers
* removes references to IndexAnalyzers from DocumentMapperParser and TypeParser.ParserContext.
It can always be retrieve it from MapperService directly in those cases
It is important that resync actions are not rejected on the primary even if its
`write` threadpool is overloaded. Today we do this by exposing
`registerRequestHandlers` to subclasses and overriding it in
`TransportResyncReplicationAction`. This isn't ideal because it obscures the
difference between this action and other replication actions, and also might
allow subclasses to try and use some state before they are properly
initialised. This change replaces this override with a constructor parameter to
solve these issues.
Relates #40706
This commit deprecates versions of Java prior to Java 11. This commit
will cause a warning to be printed to standard error when any command
line tool is invoked, or when Elasticsearch is started. Additionally, we
log a deprecation message when Elasticsearch is started.
`TransportReplicationAction` is a rather complex beast, and some of its
concrete implementations do not need all of its features. More specifically, it
(a) chases a primary around the cluster until it manages to pin it down and
then (b) executes an action on that primary and all its replicas. There are
some actions that are coordinated by the primary itself, meaning that there is
no need for the chase-the-primary phases, and in the case of peer recovery
retention leases and primary/replica resync it is important to bypass these
first phases.
This commit is a step towards separating the `TransportReplicationAction` into
these two parts. It is a mostly mechanical sequence of steps to remove some
abstractions that are no longer in use.
Completion and DocStats are pulled from internal readers
instead of external since #33835 and #33847 which doesn't require
us to refresh after a stats call since refreshes will happen internally
anyhow and that will cause updated stats on ongoing indexing.
In 6.7.0 (#39378) we added a build type of DOCKER for the docker images, but
unfortunately earlier versions do not understand this and will reject any
transport messages that mention this build type.
This commit fixes this by reporting TAR instead of DOCKER when talking to older
nodes.
Relates (but does not fix) #40511
Relates #39378
In order to remain compatible with the existing joda based
implementation the parsing of milliseconds should support parsing single
digits instead of relying on three, even with strict formats.
This adds a few tests to duel against the existing joda based
implementation in order to ensure the parsing behaviour is the same.
Closes#40403
Currently, if Manifest write is unsuccessful (i.e. WriteStateException
is thrown) we perform cleanup of newly created metadata files.
However, this is wrong.
Consider the following sequence (caught by CI here
https://github.com/elastic/elasticsearch/issues/39077):
- cluster global data is written **successful**
- the associated manifest write **fails** (during the fsync, ie files
have been written)
- deleting (revert) the manifest files, **fails**, metadata is
therefore persisted
- deleting (revert) the cluster global data is **successful**
In this case, when trying to load metadata (after node restart
because of dirty WriteStateException), the following exception will
happen
```
java.io.IOException: failed to find global metadata [generation: 0]
```
because the manifest file is referencing missing global metadata file.
This commit checks if thrown WriteStateException is dirty and if its
we don't perform any cleanup, because new Manifest file might be
created, but its deletion has failed.
In the future, we might add more fine-grained check - perform the
clean up if WriteStateException is dirty, but Manifest deletion is
successful.
Closes https://github.com/elastic/elasticsearch/issues/39077
(cherry picked from commit 1fac56916bb3c4f3333c639e59188dbe743e385b)
On mapping updates the `text` field mapper does not update
the field types for the underlying prefix and phrase fields.
In practice this shouldn't be considered as a bug but we have
an assert in the code that check that field types in the mapper service
are identical to the ones present in field mappers.
When geo point parsing threw a parse exception, it did not consume
remaining tokens from the parser. This in turn meant that
indexing documents with malformed geo points into mappings with
ignore_malformed=true would fail in some cases, since DocumentParser
expects geo_point parsing to end on the END_OBJECT token.
Related to #17617
As part of #40177 we have added top-level pipeline aggs to
`InternalAggregations`. Given that `QuerySearchResult` holds an
`InternalAggregations` instance, there is no need to keep on setting
top-level pipeline aggs separately. Top-level pipeline aggs can then
always be transported through `InternalAggregations`. Such change is
made in a backwards compatible manner.
IOException are never thrown in any of the existing pipeline aggregation
builders. Removing the throws IOException from the create method allows
to remove it also from a couple of other methods which ends up simplifying
AggregationPhase (one less catch).
The cat recovery API is incredibly useful. Yet it is missing the start
and stop time as an option from the output. This commit adds these as
options to the cat recovery API. We elect to make these not visible by
default to avoid breaking the output that users might rely on.
To make script_score query to have the same features
as function_score query, we need to add randomScore
function.
This function produces different
random scores on different index shards.
It is also able to produce random scores
based on the internal Lucene Document Ids.
Today if you try and insert a very large number like `1e9999999` into a long
field we first construct this number as a `BigDecimal`, convert this to a
`BigInteger` and then reject it because it is out of range. Unfortunately
making such a large `BigInteger` is rather expensive.
We can avoid this expense by performing a (weaker) range check on the
`BigDecimal` representation of incoming `long`s too.
Relates #26137Closes#40323
In #33062 we introduced the `cluster.remote.*.proxy` setting for proxied
connections to remote clusters, but left it deliberately undocumented since it
needed followup work so that it could work with SNI. However, since #32517 is
now closed we can add this documentation and remove the comment about its lack
of documentation.
This commit fixes an edge case in tests where search hits are empty
after the merge but some shards returned hits. This can happen if
the total number of merged hits is less than the provided `from`.
Closes#40553
It initially mentioned the type in the exception because the type used to be
required to uniquely identify a document. This is not necessary anymore given
that indices have at most one type.
`Index` interns its name and uuid. My guess is that the main goal is to avoid
having duplicate strings in the representation of the cluster state. However
I doubt it helps much given that we have many other objects in the cluster state
that we don't try to reuse, and interning has some cost. When looking into
#40263 my profiler pointed to string interning because of the `Index` object
that is created in `QueryShardContext` as one of the bottlenecks of the
`can_match` phase.
Adds the search_as_you_type field type that acts like a text field optimized
for as-you-type search completion. It creates a couple subfields that analyze
the indexed terms as shingles, against which full terms are queried, and a
prefix subfield that analyze terms as the largest shingle size used and
edge-ngrams, against which partial terms are queried
Adds a match_bool_prefix query type that creates a boolean clause of a term
query for each term except the last, for which a boolean clause with a prefix
query is created.
The match_bool_prefix query is the recommended way of querying a search as you
type field, which will boil down to term queries for each shingle of the input
text on the appropriate shingle field, and the final (possibly partial) term
as a term query on the prefix field. This field type also supports phrase and
phrase prefix queries however
This commit adds an InboundHandler to handle inbound message processing.
With this commit, this code is moved out of the TcpTransport.
Additionally, finer grained unit tests are added to ensure that the
inbound processing works as expected
Replicated closed indices can't be indexed into or searched, and therefore don't need a shard with
full indexing and search capabilities allocated. We can save on a lot of heap memory for those
indices by not allocating a mapper service and caching infrastructure (which preallocates a constant
amount per instance). Before this change, a 1GB ES instance could host 250 replicated closed
metricbeat indices (each index with one shard). After this change, the same instance can host 7300
replicated closed metricbeat instances (not that this would be a recommended configuration). Most
of the remaining memory is in the cluster state and the IndexSettings object.
Switches "discovery.type: single-node" from using a separate implementation for single-node discovery to using the existing standard discovery implementation, with two small adaptions:
- auto-bootstrapping, but requiring initial_master_nodes not to be set.
- not actively pinging other nodes using the Peerfinder
- not allowing other nodes to join its single-node cluster (if they have e.g. been set up using regular discovery and connect to the single-disco node).
Currently there are some components of message serializer and sending
that still occur in TcpTransport. This commit makes it possible to
send a message without the TcpTransport by moving all of the remaining
application logic to the OutboundHandler. Additionally, it adds unit
tests to ensure that this logic works as expected.
This test inadvertently asserts that the election occurs after a master failure
is clean. However, messy elections are a fact of life so we should not fail on
a messy election.
This change moves this test away from an `AbstractDisruptionTestCase` since it
does not need the fault detector to be so enthusiastic, and weakens the
assertions to merely say that we ignore states published by the old master
without saying anything about the cleanliness of the election.
Closes#36556
Currently the TransportMessageListener is applied and used in the
Transport class. However, local requests and responses never make it to
this class. This PR moves the listener add/remove methods to the
TransportService. After this change the Transport can only have one
listener set with it. This one listener is the TransportService, which
will then propogate the events to the external listeners.
Additionally this commit back ports #40237
Remove Tracer from MockTransportService
Currently the TransportMessageListener is applied and used in the
Transport class. However, local requests and responses never make it to
this class. This PR moves the listener add/remove methods to the
TransportService. After this change the Transport can only have one
listener set with it. This one listener is the TransportService, which
will then propogate the events to the external listeners.
Java-time fails parsing composite patterns when first pattern matches only the prefix of the input. It expects pattern in longest to shortest order. Because of this constructing just one DateTimeFormatter with appendOptional is not sufficient. Parsers have to be iterated and if the parsing fails, the next one in order should be used. In order to not degrade performance parsing should not be throw exceptions on failure. Format.parseObject was used as it only returns null when parsing failed and allows to check if full input was read.
closes#39916
backport #40100
The implementation of TransportIndexAction and TransportDeleteAction as
TransportReplicationAction existed for interoperability with older 5.x nodes, as these older nodes
coordinated single index / deletes as replication requests. This BWC layer is no longer needed in 7.x,
where these single actions are now mapped to bulk requests. Completely removing the deprecated
transport actions is not possible yet if we want to keep BWC with a 6.x transport client. The best
way here is to wait for the transport client to go away and then just remove the actions.
Each cluster state publication schedules a cancellation task with the provided publication timeout
(30s by default). This scheduled cancellation keeps a reference to the publication, and therefore the
full cluster state that was published. In case of frequently updating a large cluster state, this results
in a large number of cancellation tasks keeping references to all previously published cluster states.
FilterDirectory.getPendingDeletions does not delegate, fixed
temporarily by overriding in StoreDirectory.
This in turn caused duplicate file name use after a trimUnsafeCommits
had been done, since a new IndexWriter would not consider the pending
deletes in IndexFileDeleter. This should only happen on windows (AFAIK).
Reenabled doing index updates for all tests using
IndexShardTests.indexOnReplicaWithGaps (which could fail due to above
when using mocked WindowsFS).
Added getPendingDeletions delegation to all elasticsearch
FilterDirectory subclasses that were not trivial test-only overrides to
minimize the risk of hitting this issue in another case.
This test checks that interval queries constructed against a field with no indexed
positions will throw exceptions. It uses a randomly-build IntervalsSourceProvider
against a fixed set of fields; however, the random source builder can occasionally
provide a source with a fixed field, meaning that even if the top-level query asks
for a set of intervals over a non-indexed field, the source will delegate to another
field, and no exception will be thrown.
This commit changes the test to always use a simple Match provider.
Fixes#40436
* Log Warning on Failed Blob Deletes in BlobStoreRepository
* We should not just debug log these spots, they all can and will lead to leaked files when snapshot deletion fails
Right now, the stats API only provides refresh metrics regarding
internal refreshes. This isn't very useful and somewhat misleading for
cluster administrators since the internal refreshes are not indicative
of documents being available for search.
In this PR I added a new metric for collecting external refreshes as
they occur and exposing them through the stats API. Now, calling an
endpoint for stats will yield external refresh metrics as well.
Relates #36712
In some cases, a request to perform a retention lease action can arrive
on a primary shard before it is active. In this case, the primary shard
would not yet be in primary mode, tripping an assertion in the
replication tracker. Instead, we should not attempt to perform such
actions on an initializing shard. This commit addresses this by not
returning the primary shard in the single shard iterator if the primary
shard is not yet active.
We were accidentally not mapping the index, which meant dynamic mapping
was choosing floats for the values. This led to enough loss of precision
for the aggregated values to differ slightly from the test doubles,
which accumulated into large differences in the holt output.
This test fix adds an explicit mapping.
This commit adjusts the frequency with which CCR renews retention leases
and with which primaries sync retention leases to replicas. This helps
Lucene reclaim soft-deleted documents more aggressively, which we have
found in some use-cases can help improve performance, and either way
will help keep disk space under more control.
This is the equivalent of the `field_masking_span` query, allowing users to
merge intervals from multiple fields - for example, to search for stemmed tokens
near unstemmed tokens.
Currently, we cannot update index setting index.translog.sync_interval if index is open, because it's
not dynamic which can be updated for closed index only.
Closes#32763
A recent refactoring (#37130) where imports got mixed up (changing Lucene's
IndexNotFoundException to Elasticsearch's IndexNotFoundException) led to many warnings being
logged in case of restoring a fresh snapshot.
This change adds an option to convert a `date` field to nanoseconds resolution
and a `date_nanos` field to millisecond resolution when sorting.
The resolution of the sort can be set using the `numeric_type` option of the
field sort builder. The conversion is done at the shard level and is restricted
to dates from 1970 to 2262 for the nanoseconds resolution in order to avoid
numeric overflow.
If a replica were first reset due to one primary failover and then
promoted (before resync completes), its MSU would not include changes
since global checkpoint, leading to errors during translog replay.
Fixed by re-initializing MSU before restoring local history.
Today we don't return segments stats for closed indices which makes it
hard to tell how much memory such an index would require. With this change
we return the statistics if requested by setting `include_unloaded_segments` to
true on the rest request.
Relates to #39512
Today RareClusterStateIT#testAssignmentWithJustAddedNodes fails on my Mac
because it waits for the default connection timeout of 30 seconds to connect to
a fake node with IP address 0.0.0.0. This connection attempt fails much more
quickly on Linux so the test passes.
This commit fixes this by reducing the connection timeout for this test.
Unlike index operations which can fail at the document level to
analyzing errors, delete operations should never fail at the document
level whether soft-deletes is enabled or not. With this change, we will
always fail the engine if we fail to apply a delete operation to Lucene.
Closes#33256
We currently convert pipeline aggregators to their corresponding
InternalAggregation instance as part of the final reduction phase.
They arrive to the coordinating node as part of QuerySearchResult
objects fom the shards and, despite we may incrementally reduce
aggs (hence we may have some non-final reduce and the final
one later) all the reduction phases happen on the same node.
With CCS minimizing roundtrips though, each cluster performs its
own non-final reduction, and then serializes the results back to
the CCS coordinating node which will perform the final coordination.
This breaks the assumptions made up until now around reductions
happening all on the same node.
With #40101 we have made sure that top-level pipeline aggs are not
reduced as part of the non-final reduction. The next step is to make
sure that they don't get lost, meaning that each coordinating node
needs to send them back to the CCS coordinating node as part of
the top-level `InternalAggregations` object.
Closes#40059
Today a coordinating node forces a final reduction of sibling pipeline aggregators whenever reducing aggs, unless it is reducing aggs incrementally. This works well for incremental reduction of aggs, but breaks CCS when minimizing roundtrips as each cluster ends up reducing its own pipeline aggregators locally while that should only be done by the CCS coordinating node later. This causes issues as after their reduction, pipeline aggs cannot be further reduced, which is what happens with CCS causing errors like "java.lang.UnsupportedOperationException: Not supported" being returned.
Each coordinating node should rather honour the reduce context flag that
indicates whether we are executing a final reduce or not. If not, it should leave the sibling pipeline aggregations alone.
Note that his bug affects only pipeline aggs that don't have a parent in
the aggs tree, while all the others work well.
Relates to #40059 but does not fix it yet, as the CCS coordinating node also needs to be adapted to recreate sibling pipeline aggregators from the request.
When minimizing round-trips, each cluster returns its own independent
search response. In case sort by field and/or field collapsing were
requested, when one cluster has no results to return, the information
about the field that sorting was based on (SortField array) as well as
the field (and the values) that collapsing was performed on are missing
in the search response. That causes problems as we can't build the
proper `TopDocs` instance which would need to be either `TopFieldDocs`
or `CollapseTopFieldDocs`. The merge routine expects that all the top
docs are of the same exact type which can't be guaranteed. Given that
the problematic results are empty, hence have no impact on the final
results, we can simply skip them.
Relates to #32125Closes#40067
When using DFS_QUERY_THEN_FETCH search type, the dfs phase is run and
its results are used in the query phase to make scoring accurate.
When using CCS, depending on whether the DFS phase runs in the CCS
coordinating node (like if all shards were local) or in each remote
cluster (when minimizing round-trips), scoring will differ.
This commit disables minimizing round-trips whenever DFS is requested,
as it is not currently possible to ensure that scoring is accurate in
that case.
Relates to #32125
When a node is repurposed to master/no-data or no-master/no-data, v7.x
will not start (see #37748 and #37347). The `elasticsearch repurpose`
tool can fix this by cleaning up the problematic data.
The cache used in linearizability checker now uses approximately 6x less
memory by changing the cache from a set of (bits, state) tuples into a
map from bits -> { state }.
Each combination of states is kept once only, building on the
assumption that the number of state permutations is small compared to
the number of bits permutations. For those histories that are difficult
to check we will have many bits combinations that use the same state
permutations.
We end up now using approximately 15 bytes per entry compared to 101
bytes before, ie. a 6x improvement, allowing us to linearizability check
significantly longer histories.
Re-enabled linearizability checker in CoordinatorTests, hoping above
ensures we no longer run out of memory.
Resolves#39437
We introduced WAIT_CLUSTERSTATE action in #19287 (5.0), but then stopped
using it since #25692 (6.0). This change removes that action and related
code in 7.x and 8.0.
Relates #19287
Relates #25692
This PR introduces AsyncRecoveryTarget which executes remote calls of
peer recovery asynchronously. In this change, we also add a new
assertion to ensure that method sendBatch, which sends a batch of
history operations in phase2, is never called recursively on the same
thread. This new assertion will also be used in method sendFileChunks.
This change adds a wrapper for IndexSearcher that makes IndexSearcher#search(List, Weight, Collector) visible by
sub-classes. The wrapper is used by the ContextIndexSearcher to call this protected method on a searcher created by a plugin.
This ensures that an override of the protected method in an IndexSearcherWrapper plugin is called when a search is executed.
Closes#30758
This change adds an option to the `FieldSortBuilder` that allows to transform the type
of a numeric field into another. Possible values for this option are `long` that transforms
the source field into an integer and `double` that transforms the source field into a floating point.
This new option is useful for cross-index search when the sort field is mapped differently on some
indices. For instance if a field is mapped as a floating point in one index and as an integer in another
it is possible to align the type for both indices using the `numeric_type` option:
```
{
"sort": {
"field": "my_field",
"numeric_type": "double" <1>
}
}
```
<1> Ensure that values for this field are transformed to a floating point if needed.
This commit removes the cluster state size field from the cluster state
response, and drops the backwards compatibility layer added in 6.7.0 to
continue to support this field. As calculation of this field was
expensive and had dubious value, we have elected to remove this field.
Currently, we maintain a transport name ("mock-nio", "nio", "netty")
that is passed to a `TcpTransportChannel` when a request is received.
The value of this name is to associate with the task when we register a
task with the task manager. However, it is only possible to run ES with
one transport, so having an implementation specific name is unnecessary.
This commit removes the name and replaces it with the generic
"transport".
The `sampler` agg creates a BestDocsDeferringCollector, which internally
initializes a priority queue of size `shardSize`. This queue is
populated with empty `Object` sentinels, which is roughly 16b per
object.
Similarly, the Diversified samplers create a DiversifiedTopDocsCollectors
which internally track PQ slots with ScoreDocKeys, weighing in around
28kb
If the user sets a very abusive `shard_size`, this could easily OOM
a node or cluster since these PQ are allocated up-front without
any checks.
This commit makes sure that when we create the collector, it
cannot be larger than the maxDoc so that we don't accidentally blow
up the node. We ensure the size is not greater than the overall
index maxDoc. A similar treatment is done for `maxDocsPerValue`
parameter of the diversified samplers
For good measure, this also adds in some CB accounting to try and track
memory usage.
Finally, a redundant array creation is removed to reduce a bit of
temporary memory.
We call `ensureConnections()` to undo the effects of a disruption. However, it
is possible that one or more targets are currently CONNECTING and have been
since the disruption was active, and that the connection attempt was thwarted
by a concurrent disruption to the connection. If so, we cannot simply add our
listener to the queue because it will be notified when this CONNECTING activity
completes even though it was disrupted. We must therefore wait for all the
current activity to finish and then go through and reconnect to any missing
nodes.
Closes#40030.
Today we load the shard history retention leases from disk whenever opening the
engine, and treat a missing file as an empty set of leases. However in some
cases this is inappropriate: we might be restoring from a snapshot (if the
target index already exists then there may be leases on disk) or
force-allocating a stale primary, and in neither case does it make sense to
restore the retention leases from disk.
With this change we write an empty retention leases file during recovery,
except for the following cases:
- During peer recovery the on-disk leases may be accurate and could be needed
if the recovery target is made into a primary.
- During recovery from an existing store, as long as we are not
force-allocating a stale primary.
Relates #37165
Today we test Zen1/Zen2 compatibility by running 7.x nodes with a "fake" Zen1
implementation. However this is not a truly faithful test because these nodes
do known how to properly deserialize a 7.x cluster state, voting configurations
and all, whereas a real Zen1 node is in 6.7 and ignores the coordination
metadata.
We only ever apply a cluster state that's been committed, which in Zen2
involves setting the last-committed configuration to equal the last-accepted
configuration. Zen1 knows nothing about this adjustment, so it is possible for
these to differ. This breaks the assertion that the cluster states are equal on
all nodes after integration tests.
This commit fixes this by implementing this adjustment in Zen1 before applying
a cluster state.
Fixes#40055.
This change ensures that we do not make assumptions about the length
of the input that we can read from the stdin. It still consumes only
one line, as the previous implementation
* Handle UTF8 values in the keystore
Our current implementation uses CharBuffer#array to get the chars
that were decoded from the UTF-8 bytes. The backing array of
CharBuffer is created in CharsetDecoder#decode and gets an initial
length that is the same as the length of the ByteBuffer it decodes,
hence the number of UTF-8 bytes.
This works fine for the first 128 characters where each one needs
one bytes, but for the next UTF-8 characters (other latin alphabets
Greek, Cyrillic etc.) where we need 2 to 4 bytes per character, this
backing char array has a larger size than the number of the actual
chars this CharBuffer contains. Calling `array()` on it will return
a char array that can potentially have extra null chars so the
SecureString we get from the KeystoreWrapper, is not the same as the
one we entered.
This commit changes the behavior to use Arrays#copyOfRange to get
the necessary chars from the CharBuffer and adds a test with
random ( maybe not printable ) UTF-8 strings
Computing the compressed size of the cluster state on every invocation
of cluster:monitor/state action is expensive, and the value of this
field is dubious anyway. Therefore we want to remove computing this
field. As a first step, we stop computing and return this field by
default. To avoid breaking users, we will give them a system property to
use to tide them over until the next major release when we will actually
remove this field. This comes with a deprecation warning too, and the
backport to the appropriate minor will also include a note in the
migration guide. There will be a follow-up to remove this field in the
next major version.
After the joda-java time migration we were formatting zone ids with zoneOrOffsetId method. This when a date was provided with a ZoneRegion for instance America/Edmonton it was appending this zone identifier instead of zone formatted as +HH:MM.
This fix is changing the format of zone suffix for all printers and also always wrapping a Temporal into a ZonedDateTime when formatting.
closes#38471
backport #39568
Currently there is a method `Recycler#obtain(size)` that allows a size
parameter to be passed. However all implementations ignore this
parameter and just allocate a page size based on other settings. This
commit removes this method.
When performing the test with 57 master-eligible nodes and one node
crash, we saw messy elections, when multiple nodes were attempting to
become master.
JoinHelper has logged 105 long log messages with lengthy stack
traces during one such election.
To address this, we decided to log these messages every time only on
debug level.
We will log last unsuccessful join attempt (along with a timestamp)
if any with WARN level if the cluster is failing to form.
(cherry picked from commit 17a148cc27b5ac6c2e04ef5ae344da05a8a90902)
Currently token filter settings are treated as fixed once they are declared and
used in an analyzer. This is done to prevent changes in analyzers that are already
used actively to index documents, since changes to the analysis chain could
corrupt the index. However, it would be safe to allow updates to token
filters at search time ("search_analyzer"). This change introduces a new
property of token filters that allows to mark them as only being usable at search
or at index time. Any analyzer that uses these tokenfilters inherits that property
and can be rejected if they are used in other contexts. This is a first step towards
making specific token filters (e.g. synonym filter) updateable.
Relates to #29051
Today, when applying new cluster state we attempt to connect to all of its
nodes as a blocking part of the application process. This is the right thing to
do with new nodes, and is a no-op on any already-connected nodes, but is
questionable on known nodes from which we are currently disconnected: there is
a risk that we are partitioned from these nodes so that any attempt to connect
to them will hang until it times out. This can dramatically slow down the
application of new cluster states which hinders the recovery of the cluster
during certain kinds of partition.
If nodes are disconnected from the master then it is likely that they are to be
removed as part of a subsequent cluster state update, so there's no need to try
and reconnect to them like this. Moreover there is no need to attempt to
reconnect to disconnected nodes as part of the cluster state application
process, because we periodically try and reconnect to any disconnected nodes,
and handle their disconnectedness reasonably gracefully in the meantime.
This commit alters this behaviour to avoid reconnecting to known nodes during
cluster state application.
Resolves#29025.
If a primary on 6.7 and a replica on 5.6 are running more than 5 minutes
(retention leases background sync interval), the retention leases
background sync will be triggered, and it will trip 6.7 node due to the
illegal checkpoint value. We can fix the problem by making the returned
checkpoint depends on the node version. This PR, however, chooses to
enforce retention leases require soft deletes, and make retention leases
sync noop if soft deletes is disabled instead.
Closes#39914
* The test failure in #39852 is caused by a file in the initial repository when there should not be any
* It seems that on a normal consistent file system no left-over file should exist ever here after the validation finishes and I can't reproduce or see any other path to a dangling file in the fresh respository
=> added a more verbose and strict assertion that will log what file is left over next time
* Relates #39852
This commit removes the "doc" type from monitoring internal indexes.
The template still carries the "_doc" type since that is needed for
the internal representation.
This change impacts the following templates:
monitoring-alerts.json
monitoring-beats.json
monitoring-es.json
monitoring-kibana.json
monitoring-logstash.json
As part of the required changes, the system_api_version has been
bumped from "6" to "7" and support for version "2" has been dropped.
A new empty pipeline is now introduced for the version "7", and
the formerly empty "6" pipeline will now remove the type and re-direct
the request to the "7" index.
Additionally, to due to a difference in the internal representation
(which requires the inclusion of "_doc" type) and external representation
(which requires the exclusion of any type) a helper method is introduced
to help convert internal to external representation, and used by the
monitoring HTTP template exporter.
Relates #38637
Executors of type fixed_auto_queue_size (i.e. search / search_throttled) wrap runnables into
TimedRunnable, which is an AbstractRunnable. This is dangerous as it might silently swallow
exceptions, and possibly miss calling a response listener. While this has not triggered any failures in
the tests I have run so far, it might help uncover future problems.
Follow-up to #36137
This test started failing since decreasing the leader and follower check timeouts (#38298). The
reason is that the test was relying on the default publication timeout to come into effect before
leader / follower check timeouts, which is now not always true anymore.
Closes#38867
Nhat Nguyen
Michael Morello
Nick Knize
Armin Braun
Christoph Büscher
Jim Ferenczi
Ignacio Vera
Luca Cavanna
Przemyslaw Gomulka
Zachary Tong
Jason Tedor
David Roberts
David Turner
Jim Ferenczi
Simon Willnauer
Adrien Grand
Alpar Torok
Yogesh Gaikwad
Tim Brooks
Ignacio Vera
Antonio Matarrese
Ryan Ernst
Zachary Tong
Dimitrios Liappis
Julie Tibshirani
Mark Vieira
Henning Andersen
Colin Goodheart-Smithe
Martijn van Groningen
Mayya Sharipova
Yannick Welsch
Alexander Reelsen
Andrey Ershov
Oghenovo Usiwoma
Andy Bristol
Benjamin Trent
alex101101
Daniel Mitterdorfer
Alan Woodward
Like
Dimitris Athanasiou
Albert Zaharovits
Ioannis Kakavas
Tamara Braun
lzh3636
Jake Landis