The following edge cases were fixed:
1. A request to force-stop a stopping datafeed is no longer
ignored. Force-stop is an important recovery mechanism
if normal stop doesn't work for some reason, and needs
to operate on a datafeed in any state other than stopped.
2. If the node that a datafeed is running on is removed from
the cluster during a normal stop then the stop request is
retried (and will likely succeed on this retry by simply
cancelling the persistent task for the affected datafeed).
3. If there are multiple simultaneous force-stop requests for
the same datafeed we no longer fail the one that is
processed second. The previous behaviour was wrong as
stopping a stopped datafeed is not an error, so stopping
a datafeed twice simultaneously should not be either.
Backport of #49191
* [ML] ML Model Inference Ingest Processor (#49052)
* [ML][Inference] adds lazy model loader and inference (#47410)
This adds a couple of things:
- A model loader service that is accessible via transport calls. This service will load in models and cache them. They will stay loaded until a processor no longer references them
- A Model class and its first sub-class LocalModel. Used to cache model information and run inference.
- Transport action and handler for requests to infer against a local model
Related Feature PRs:
* [ML][Inference] Adjust inference configuration option API (#47812)
* [ML][Inference] adds logistic_regression output aggregator (#48075)
* [ML][Inference] Adding read/del trained models (#47882)
* [ML][Inference] Adding inference ingest processor (#47859)
* [ML][Inference] fixing classification inference for ensemble (#48463)
* [ML][Inference] Adding model memory estimations (#48323)
* [ML][Inference] adding more options to inference processor (#48545)
* [ML][Inference] handle string values better in feature extraction (#48584)
* [ML][Inference] Adding _stats endpoint for inference (#48492)
* [ML][Inference] add inference processors and trained models to usage (#47869)
* [ML][Inference] add new flag for optionally including model definition (#48718)
* [ML][Inference] adding license checks (#49056)
* [ML][Inference] Adding memory and compute estimates to inference (#48955)
* fixing version of indexed docs for model inference
The logic for `cleanupInProgress()` was backwards everywhere (method itself and
all but one user). Also, we weren't checking it when removing a repository.
This lead to a bug (in the one spot that didn't use the method backwards) that prevented
the cleanup cluster state entry from ever being removed from the cluster state if master
failed over during the cleanup process.
This change corrects the backwards logic, adds a test that makes sure the cleanup
is always removed and adds a check that prevents repository removal during cleanup
to the repositories service.
Also, the failure handling logic in the cleanup action was broken. Repeated invocation would lead to the cleanup being removed from the cluster state even if it was in progress. Fixed by adding a flag that indicates whether or not any removal of the cleanup task from the cluster state must be executed. Sorry for mixing this in here, but I had to fix it in the same PR, as the first test (for master-failover) otherwise would often just delete the blocked cleanup action as a result of a transport master action retry.
improve error handling for script errors, treating it as irrecoverable errors which puts the task
immediately into failed state, also improves the error extraction to properly report the script
error.
fixes#48467
AutoFollowIT relies on assertBusy() calls to wait for a given number of
leader indices to be created but this is prone to failures on CI. Instead,
we should use latches to indicate when auto-follow patterns must be
paused and resumed.
This commit fixes a NPE problem as reported in #49150.
But this problem uncovered that we never added proper handling
of state for data frame analytics tasks.
In this commit we improve the `MlTasks.getDataFrameAnalyticsState`
method to handle null tasks and state tasks properly.
Closes#49150
Backport of #49186
If CCR encounters a rejected execution exception, today we treat this as
fatal. This is not though, as the stuffed queue could drain. Requiring
an administrator to manually restart the follow tasks that faced such an
exception is a burden. This commit addresses this by making CCR
auto-retry on rejected execution exceptions.
We can't guarantee expected request failures if search request is across
many indexes, as if expected shards fail, some indexes may return 200.
closes#47743
The JdbcHttpClientRequestTests and HttpClientRequestTests classes both
hold a static reference to a mock web server that internally uses the
JDKs built-in HttpServer, which resides in a sun package that the
RamUsageEstimator does not have access to. This causes builds that use
a runtime of Java 8 to fail since the StaticFieldsInvariantRule is run
when Java 8 is used.
Relates #41526
Relates #49105
When triggered either by becoming master, a new cluster state, or a
periodic schedule, an ILM policy execution through
`maybeRunAsyncAction`, `runPolicyAfterStateChange`, or
`runPeriodicStep` throwing an exception will cause the loop the
terminate. This means that any indices that would have been processed
after the index where the exception was thrown will not be processed by
ILM.
For most execution this is not a problem because the actual running of
steps is protected by a try/catch that moves the index to the ERROR step
in the event of a problem. If an exception occurs prior to step
execution (for example, in fetching and parsing the current
policy/step) however, it causes the loop termination previously
mentioned.
This commit wraps the invocation of the methods specified above in a
try/catch block that provides better logging and does not bubble the
exception up.
Backport of #47468 to 7.x
This PR adds a new metric aggregation called string_stats that operates on string terms of a document and returns the following:
min_length: The length of the shortest term
max_length: The length of the longest term
avg_length: The average length of all terms
distribution: The probability distribution of all characters appearing in all terms
entropy: The total Shannon entropy value calculated for all terms
This aggregation has been implemented as an analytics plugin.
The rollover action is now a retryable step (see #48256)
so ILM will keep retrying until it succeeds as opposed to stopping and
moving the execution in the ERROR step.
Fixes#49073
(cherry picked from commit 3ae90898121b43032ec8f3b50514d93a86e14d0f)
Signed-off-by: Andrei Dan <andrei.dan@elastic.co>
# Conflicts:
# x-pack/plugin/ilm/qa/multi-node/src/test/java/org/elasticsearch/xpack/ilm/TimeSeriesLifecycleActionsIT.java
Ensures that methods that are called from different threads ( i.e.
from the callbacks of org.apache.http.concurrent.FutureCallback )
catch `Exception` instead of only the expected checked exceptions.
This resolves a bug where OpenIdConnectAuthenticator#mergeObjects
would throw an IllegalStateException that was never caught causing
the thread to hang and the listener to never be called. This would
in turn cause Kibana requests to authenticate with OpenID Connect
to timeout and fail without even logging anything relevant.
This also guards against unexpected Exceptions that might be thrown
by invoked library methods while performing the necessary operations
in these callbacks.
The cluster state is obtained twice in the EnrichPolicyRunner when updating
the final alias. There is a possibility for the state to be slightly different
between those two calls. This PR just has the function get the cluster state
once and reuse it for the life of the function call.
Backport of #48849. Update `.editorconfig` to make the Java settings the
default for all files, and then apply a 2-space indent to all `*.gradle`
files. Then reformat all the files.
Temporarily "mute" the testReplaceChildren for Pivot since it leads to
failing tests for some seeds, since the new child doesn't respond to a
valid data type.
Relates to #48900
(cherry picked from commit 6200a2207b9a4264d2f3fc976577323c7e084317)
We can have a race here where `scheduleNextRun` executes concurrently to `stop`
and so we run into a `RejectedExecutionException` that we don't catch and thus it
fails tests.
=> Fixed by ignoring these so long as they coincide with a scheduler shutdown
When a node shuts down, `TransportService` moves to stopped state and
then closes connections. If a request is done in between, an exception
was thrown that was not retried in replication actions. Now throw a
wrapped `NodeClosedException` exception instead, which is correctly
handled in replication action. Fixed other usages too.
Relates #42612
This commit fixes an off-by-one bug in the AutoFollowIT test that causes
failures because the leaderIndices counter is incremented during the evaluation
of the leaderIndices.incrementAndGet() < 20 condition but the 20th index is
not created, making the final assertion not verified.
It also gives a bit more time for cluster state updates to be processed on the
follower cluster.
Closes#48982
Ensures that we always use the primary term established by the primary to index docs on the
replica. Makes the logic around replication less brittle by always using the operation primary
term on the replica that is coming from the primary.
Our documentation regarding FIPS 140 claimed that when using SAML
in a JVM that is configured in FIPS approved only mode, one could
not use encrypted assertions. This stemmed from a wrong
understanding regarding the compliance of RSA-OAEP which is used
as the key wrapping algorithm for encrypting the key with which the
SAML Assertion is encrypted.
However, as stated for instance in
https://downloads.bouncycastle.org/fips-java/BC-FJA-SecurityPolicy-1.0.0.pdf
RSA-OAEP is approved for key transport, so this limitation is not
effective.
This change removes the limitation from our FIPS 140 related
documentation.
This PR makes the following two fixes around updating flattened fields:
* Make sure that the new value for ignore_above is immediately taken into
affect. Previously we recorded the new value but did not use it when parsing
documents.
* Allow depth_limit to be updated dynamically. It seems plausible that a user
might want to tweak this setting as they encounter more data.
When using the move-to-step API, we should reread the phase JSON from
the latest version of the ILM policy. This allows a user to move to the
same step while re-reading the policy's latest version. For example,
when changing rollover criteria.
While manually messing around with some other things I discovered that
we only reread the policy when using the retry API, not the move-to-step
API. This commit changes the move-to-step API to always read the latest
version of the policy.
Backport of #48908
The enrich project doesn't have much history as all the other gradle projects,
so it makes sense to enable spotless for this gradle project.
The `HttpExportBulk` exporter is using a lot more memory than it needs to
by allocating buffers for serialization and IO:
* Remove copying of all bytes when flushing, instead use the stream wrapper
* Remove copying step turning the BAOS into a `byte[]`
* This also avoids the allocation of a single huge `byte[]` and instead makes use of the internal paging logic of the `BytesStreamOutput`
* Don't allocate a new BAOS for every document, just keep appending to a single BAOS
Currently the BulkProcessor class uses a single scheduler to schedule
flushes and retries. Functionally these are very different concerns but
can result in a dead lock. Specifically, the single shared scheduler
can kick off a flush task, which only finishes it's task when the bulk
that is being flushed finishes. If (for what ever reason), any items in
that bulk fails it will (by default) schedule a retry. However, that retry
will never run it's task, since the flush task is consuming the 1 and
only thread available from the shared scheduler.
Since the BulkProcessor is mostly client based code, the client can
provide their own scheduler. As-is the scheduler would require
at minimum 2 worker threads to avoid the potential deadlock. Since the
number of threads is a configuration option in the scheduler, the code
can not enforce this 2 worker rule until runtime. For this reason this
commit splits the single task scheduler into 2 schedulers. This eliminates
the potential for the flush task to block the retry task and removes this
deadlock scenario.
This commit also deprecates the Java APIs that presume a single scheduler,
and updates any internal code to no longer use those APIs.
Fixes#47599
Note - #41451 fixed the general case where a bulk fails and is retried
that can result in a deadlock. This fix should address that case as well as
the case when a bulk failure *from the flush* needs to be retried.
* [ML] Add new geo_results.(actual_point|typical_point) fields for `lat_long` results (#47050)
[ML] Add new geo_results.(actual_point|typical_point) fields for `lat_long` results (#47050)
Related PR: https://github.com/elastic/ml-cpp/pull/809
* adjusting bwc version
The timeout was increased to 60s to allow this test more time to reach a
yellow state. However, the test will still on occasion fail even with the
60s timeout.
Related: #48381
Related: #48434
Related: #47950
Related: #40178
CCR follower stats can return information for persistent tasks that are in the process of being cleaned up. This is problematic for tests where CCR follower indices have been deleted, but their persistent follower task is only cleaned up asynchronously afterwards. If one of the following tests then accesses the follower stats, it might still get the stats for that follower task.
In addition, some tests were not cleaning up their auto-follow patterns, leaving orphaned patterns behind. Other tests cleaned up their auto-follow patterns. As always the same name was used, it just depended on the test execution order whether this led to a failure or not. This commit fixes the offensive tests, and will also automatically remove auto-follow-patterns at the end of tests, like we do for many other features.
Closes #48700
We might have some outstanding renew retention lease requests after a
shard has unfollowed. If testRetentionLeaseIsAddedIfItDisappearsWhileFollowing
intercepts a renew request from other tests then we will never unlatch
and the test will time out.
Closes#45192
The loading of `RepositoryData` is not an atomic operation.
It uses a list + get combination of calls.
This lead to accidentally returning an empty repository data
for generations >=0 which can never not exist unless the repository
is corrupted.
In the test #48122 (and other SLM tests) there was a low chance of
running into this concurrent modification scenario and the repository
actually moving two index generations between listing out the
index-N and loading the latest version of it. Since we only keep
two index-N around at a time this lead to unexpectedly absent
snapshots in status APIs.
Fixing the behavior to be more resilient is non-trivial but in the works.
For now I think we should simply throw in this scenario. This will also
help prevent corruption in the unlikely event but possible of running into this
issue in a snapshot create or delete operation on master failover on a
repository like S3 which doesn't have the "no overwrites" protection on
writing a new index-N.
Fixes#48122
testRetentionLeaseIsAddedIfItDisappearsWhileFollowing is still failing
although we already have several fixes. I think other tests interfere
and cause this test to fail. We can use the test scope to isolate them.
However, I prefer to add debug logs so we can find the source.
Relates #45192
Previously this step moved to the forcemerge step, however, if the
machine running the test was fast enough, it would execute the
forcemerge and move to the next step (`segment-count`) so the comparison
would fail. This commit changes the step to be a step that will never go
anywhere else, the terminal step.
Resolves#48761
decouple TransformTask and ClientTransformIndexer. Interaction between the 2 classes are
now moved into a context class which holds shared information.
relates #45369
This commit introduces a consistent, and type-safe manner for handling
global build parameters through out our build logic. Primarily this
replaces the existing usages of extra properties with static accessors.
It also introduces and explicit API for initialization and mutation of
any such parameters, as well as better error handling for uninitialized
or eager access of parameter values.
Closes#42042
In the case multi-fields exist in the source index, we pick
all variants of them in our extracted fields detection for
data frame analytics. This means we may have multiple instances
of the same feature. The worse consequence of this is when the
dependent variable (for regression or classification) is also
duplicated which means we train a model on the dependent variable
itself.
Now that #48770 is merged, this commit is adding logic to
only select one variant of multi-fields.
Closes#48756
Backport of #48799
* ILM Test asserts on the same ilm/_explain output
With the introduction of retryable steps subsequent ilm/_explain calls
can see the state of an ilm cycle move out of the error step. This test
made several assertions assuming that the cycle remains in the error
step so this commit changes the test to make one _explain call and have
all the asserts work on the same ilm state (so subsequent assumptions to
the cycle being in the error step are valid).
* Drop unused field in test.
(cherry picked from commit 44c74bb487151c886a08b27f32b13f7a72056997)
Signed-off-by: Andrei Dan <andrei.dan@elastic.co>
Aggregatable mutli-fields are at the moment wrongly mapped
as normal doc_value fields and thus they support fetching from
source. However, they do not exist in the source. This results
to failure to extract such fields.
This commit fixes this bug. While a fix could be worked out
on top of the existing code, it is evident the extraction logic
has become difficult to understand and maintain. As we also
want to deduplicate multi-fields for data frame analytics,
it seemed appropriate to refactor the code to simplify and
better handle the extraction of multi-fields.
Relates #48756
Backport of #48770
* Introduce binary_format request parameter (binary.format for JDBC) to disable binary
communication between clients (jdbc/odbc) and server.
* for CLI - "binary" command line parameter (or -b) is introduced. Default value is "true".
* binary communication (cbor) is enabled by default
* disabling request parameter introduced for debugging purposes only
(cherry picked from commit f96a5ca61cb9fad9ed59357320af20e669348ce7)
* Add ingest info to Cluster Stats (#48485)
This commit enhances the ClusterStatsNodes response to include global
processor usage stats on a per-processor basis.
example output:
```
...
"processor_stats": {
"gsub": {
"count": 0,
"failed": 0
"current": 0
"time_in_millis": 0
},
"script": {
"count": 0,
"failed": 0
"current": 0,
"time_in_millis": 0
}
}
...
```
The purpose for this enhancement is to make it easier to collect stats on how specific processors are being used across the cluster beyond the current per-node usage statistics that currently exist in node stats.
Closes#46146.
* fix BWC of ingest stats
The introduction of processor types into IngestStats had a bug.
It was set to `null` and set as the key to the map. This would
throw a NPE. This commit resolves this by setting all the processor
types from previous versions that are not serializing it out to
`_NOT_AVAILABLE`.
This test used an index without an alias to simulate a failure in the
`check-rollover-ready` step. However, with #48256 that step
automatically retries, meaning that the index may not always be in
the ERROR step.
This commit changes the test to use a shrink action with an invalid
number of shards so that it stays in the ERROR step.
Resolves#48767
Previous behavior while copying HTTP headers to the ThreadContext,
would allow multiple HTTP headers with the same name, handling only
the first occurrence and disregarding the rest of the values. This
can be confusing when dealing with multiple Headers as it is not
obvious which value is read and which ones are silently dropped.
According to RFC-7230, a client must not send multiple header fields
with the same field name in a HTTP message, unless the entire field
value for this header is defined as a comma separated list or this
specific header is a well-known exception.
This commits changes the behavior in order to be more compliant to
the aforementioned RFC by requiring the classes that implement
ActionPlugin to declare if a header can be multi-valued or not when
registering this header to be copied over to the ThreadContext in
ActionPlugin#getRestHeaders.
If the header is allowed to be multivalued, then all such headers
are read from the HTTP request and their values get concatenated in
a comma-separated string.
If the header is not allowed to be multivalued, and the HTTP
request contains multiple such Headers with different values, the
request is rejected with a 400 status.
Fix an issue that arises from the use of ExpressionIds as keys in a lookup map
that helps the QueryTranslator to identify the grouping columns. The issue is
that the same expression in different parts of the query (SELECT clause and GROUP BY clause)
ends up with different ExpressionIds so the lookup fails. So, instead of ExpressionIds
use the hashCode() of NamedExpression.
Fixes: #41159Fixes: #40001Fixes: #40240Fixes: #33361Fixes: #46316Fixes: #36074Fixes: #34543Fixes: #37044Fixes: #42041
(cherry picked from commit 3c38ea555984fcd2c6bf9e39d0f47a01b09e7c48)
This adds the infrastructure to be able to retry the execution of retryable
steps and makes the `check-rollover-ready` retryable as an initial step to
make the rollover action more resilient to transient errors.
(cherry picked from commit 454020ac8acb147eae97acb4ccd6fb470d1e5f48)
Signed-off-by: Andrei Dan <andrei.dan@elastic.co>
At present the ML C++ artifact is always downloaded from
S3. This change adds an option to configure the location.
(The intention is to use a file:/// URL to pick up the
artifact built in a Docker container in ml-cpp PR builds
so that C++ changes that will break Java integration tests
can be detected before the ml-cpp PRs are merged.)
Relates elastic/ml-cpp#766
Moves common field extraction logic to its own package so that it can
be used both for anomaly detection and data frame analytics.
In preparation for refactoring extraction fields to be simpler and to
support multi-fields properly.
Backport of #48709
When we load a JSON Web Key (JWKSet) from the specified
file using JWKSet.load it internally uses IOUtils.readFileToString
but the opened FileInputStream is never closed after usage.
https://bitbucket.org/connect2id/nimbus-jose-jwt/issues/342
This commit reads the file and parses the JWKSet from the string.
This also fixes an issue wherein if the underlying file changed,
for every change event it would add another file watcher. The
change is to only add the file watcher at the start.
Closes#44942
* Un-AwaitsFix and enhance logging for testPolicyCRUD
This removes the `AwaitsFix` and increases the test logging for
`SnapshotLifecycleServiceTests.testPolicyCRUD` in an effort to track
down the cause of #44997.
* Remove unused import
This PR performs the following changes:
* Split `ScoreScriptUtilsTests` into `DenseVectorFunctionTests` and
`SparseVectorFunctionTests`. This will make it easier to delete all sparse
vector function tests once we remove support on 8.x.
* As much as possible, break up the large test methods into individual tests
for each vector function (`cosineSimilarity`, `l2norm`, etc.).
The logger was erroneously using the `SnapshotLifecycleMetadata` class
for its initialization, making it hard to target packages for logging
levels since `SnapshotLifecycleMetadata` is in a different package.
* [ML][Inference] separating definition and config object storage (#48651)
This separates out the `definition` object from being stored within the configuration object in the index.
This allows us to gather the config object without decompressing a potentially large definition.
Additionally, `input` is moved to the TrainedModelConfig object and out of the definition. This is so the trained input fields are accessible outside the potentially large model definition.
This adds a guard for the SLM lifecycle and retention service that
prevents new jobs from being scheduled once the service has been
stopped. Previous if the node were shut down the service would be
stopped, but a cluster state or local master election would cause a job
to attempt to be scheduled. This could lead to an uncaught
`RejectedExecutionException`.
Resolves#47749
The 1MB IO-buffer size per transport thread is causing trouble in
some tests, albeit at a low rate. Reducing the number of transport
threads was not enough to fully fix this situation.
Allowing to configure the size of the buffer and reducing it by
more than an order of magnitude should fix these tests.
Closes#46803
Previously the functions accepted a doc values reference, whereas they now
accept the name of the vector field. Here's an example of how a vector function
was called before and after the change.
```
Before: cosineSimilarity(params.query_vector, doc['field'])
After: cosineSimilarity(params.query_vector, 'field')
```
This seems more intuitive, since we don't allow direct access to vector doc
values and the the meaning of `doc['field']` is unclear.
The PR makes the following changes (broken into distinct commits):
* Add new function signatures of the form `function(params.query_vector,
'field')` and deprecates the old ones. Because Painless doesn't allow two
methods with the same name and number of arguments, we allow a generic `Object`
to be passed in to the function and decide on the behavior through an
`instanceof` check.
* Refactor the class bindings so that the document field is passed to the
constructor instead of the instance method. This allows us to avoid retrieving
the vector doc values on every function invocation, which gives a tiny speed-up
in benchmarks.
Note that this PR adds new signatures for the sparse vector functions too, even
though sparse vectors are deprecated. It seemed simplest to understand (for both
us and users) to keep everything symmetric between dense and sparse vectors.
The format returned by the API is not always parsable with
`Instant.parse()`, so this commit adjusts to parsing those dates as
`ISO_ZONED_DATE_TIME` instead, which appears to always parse the
returned value correctly.
The failures in these tests have been remarkably difficult to track
down, in part because they will not reproduce locally. This commit
unmutes the flaky tests and increases logging, as well as introducing
some additional logging, to attempt to pin down the failures.
The open and close follower steps didn't check if the index is open,
closed respectively, before executing the open/close request.
This changes the steps to check the index state and only perform the
open/close operation if the index is not already open/closed.
The cron schedule "1 2 3 4 5 ?" will run every May 4 at 03:02:01, which
may result in unnecessary test failures once a year. This commit
switches out uses of that schedule in tests for one which will never
execute (because it specifies a day which doesn't exist, Feb. 31). Also
factors the schedule out to a constant to make the intent clearer.
I think the problem is that the master is trying to relocate the "upgraded_scroll" shard back to
the node on which it was previously allocated, but to which it can't be allocated now due to the
shard lock being held because of an in-progress scroll. As the master keeps on retrying and
retrying (and indefinitely tries so because max_retries does not apply to relocations, it blocks
any other lower-prioritized task from completing, which leads to the rolling upgrade tests failing
(see #48395).
Closes#48395
This commit simplifies and standardizes our usage of the Gradle Shadow
plugin to conform more to plugin conventions. The custom "bundle" plugin
has been removed as it's not necessary and performs the same function
as the Shadow plugin's default behavior with existing configurations.
Additionally, this removes unnecessary creation of a "nodeps" artifact,
which is unnecessary because by default project dependencies will in
fact use the non-shadowed JAR unless explicitly depending on the
"shadow" configuration.
Finally, we've cleaned up the logic used for unit testing, so we are
now correctly testing against the shadow JAR when the plugin is applied.
This better represents a real-world scenario for consumers and provides
better test coverage for incorrectly declared dependencies.
(cherry picked from commit 3698131109c7e78bdd3a3340707e1c7b4740d310)
Due to a bug, GETing a snapshot can cause a RespositoryException to be
thrown. This error is transient and should be retried, rather than
causing the test to fail. This commit converts those
RepositoryExceptions into AssertionErrors so that they will be retried
in code wrapped in assertBusy.
Benjamin Trent
Przemysław Witek
David Roberts
Lisa Cawley
Julie Tibshirani
Armin Braun
Hendrik Muhs
Tanguy Leroux
Dimitris Athanasiou
Jason Tedor
Mayya Sharipova
Jay Modi
Lee Hinman
Albert Zaharovits
Christos Soulios
Andrei Dan
Ioannis Kakavas
James Baiera
Dan Hermann
Rory Hunter
Marios Trivyzas
Henning Andersen
Yannick Welsch
Martijn van Groningen
Jake Landis
Christoph Büscher
James Rodewig
Nhat Nguyen
Mark Vieira
Tim Vernum
Andrei Stefan
Andrei Stefan
Tal Levy
Andrey Ershov
Alexander Reelsen
emasab
Albert Zaharovits
Yogesh Gaikwad
Gordon Brown
Cris da Rocha