With node ordinals gone, there's no longer a need for such a complicated full cluster restart
procedure (as we can now uniquely associate nodes to data folders).
Follow-up to #41652
Follow up to #49729
This change removes falling back to listing out the repository contents to find the latest `index-N` in write-mounted blob store repositories.
This saves 2-3 list operations on each snapshot create and delete operation. Also it makes all the snapshot status APIs cheaper (and faster) by saving one list operation there as well in many cases.
This removes the resiliency to concurrent modifications of the repository as a result and puts a repository in a `corrupted` state in case loading `RepositoryData` failed from the assumed generation.
* Remove BlobContainer Tests against Mocks
Removing all these weird mocks as asked for by #30424.
All these tests are now part of real repository ITs and otherwise left unchanged if they had
independent tests that didn't call the `createBlobStore` method previously.
The HDFS tests also get added coverage as a side-effect because they did not have an implementation
of the abstract repository ITs.
Closes#30424
Since 7.4, we switch from translog to Lucene as the source of history
for peer recoveries. However, we reduce the likelihood of
operation-based recoveries when performing a full cluster restart from
pre-7.4 because existing copies do not have PPRL.
To remedy this issue, we fallback using translog in peer recoveries if
the recovering replica does not have a peer recovery retention lease,
and the replication group hasn't fully migrated to PRRL.
Relates #45136
Today we do not use retention leases in peer recovery for closed indices
because we can't sync retention leases on closed indices. This change
allows that ability and adjusts peer recovery to use retention leases
for all indices with soft-deletes enabled.
Relates #45136
Co-authored-by: David Turner <david.turner@elastic.co>
See discussion in #50047 (comment).
There are reproducible issues with Thread#suspend in Jdk11 and Jdk12 for me locally
and we have one failure for each on CI.
Jdk8 and Jdk13 are stable though on CI and in my testing so I'd selectively disable
this test here to keep the coverage. We aren't using suspend in production code so the
JDK bug behind this does not affect us.
Closes#50047
* Remove Unused Single Delete in BlobStoreRepository
There are no more production uses of the non-bulk delete or the delete that throws
on missing so this commit removes both these methods.
Only the bulk delete logic remains. Where the bulk delete was derived from single deletes,
the single delete code was inlined into the bulk delete method.
Where single delete was used in tests it was replaced by bulk deleting.
Multiple version ranges are allowed to be used in section skip in yml
tests. This is useful when a bugfix was backported to latest versions
and all previous releases contain a wire breaking bug.
examples:
6.1.0 - 6.3.0, 6.6.0 - 6.7.9, 7.0 -
- 7.2, 8.0.0 -
backport #50014
Step on the road to #49060.
This commit adds the logic to keep track of a repository's generation
across repository operations. See changes to package level Javadoc for the concrete changes in the distributed state machine.
It updates the write side of new repository generations to be fully consistent via the cluster state. With this change, no `index-N` will be overwritten for the same repository ever. So eventual consistency issues around conflicting updates to the same `index-N` are not a possibility any longer.
With this change the read side will still use listing of repository contents instead of relying solely on the cluster state contents.
The logic for that will be introduced in #49060. This retains the ability to externally delete the contents of a repository and continue using it afterwards for the time being. In #49060 the use of listing to determine the repository generation will be removed in all cases (except for full-cluster restart) as the last step in this effort.
In order to cache script results in the query shard cache, we need to
check if scripts are deterministic. This change adds a default method
to the script factories, `isResultDeterministic() -> false` which is
used by the `QueryShardContext`.
Script results were never cached and that does not change here. Future
changes will implement this method based on whether the results of the
scripts are deterministic or not and therefore cacheable.
Refs: #49466
**Backport**
Historically only two things happened in the final reduction:
empty buckets were filled, and pipeline aggs were reduced (since it
was the final reduction, this was safe). Usage of the final reduction
is growing however. Auto-date-histo might need to perform
many reductions on final-reduce to merge down buckets, CCS
may need to side-step the final reduction if sending to a
different cluster, etc
Having pipelines generate their output in the final reduce was
convenient, but is becoming increasingly difficult to manage
as the rest of the agg framework advances.
This commit decouples pipeline aggs from the final reduction by
introducing a new "top level" reduce, which should be called
at the beginning of the reduce cycle (e.g. from the SearchPhaseController).
This will only reduce pipeline aggs on the final reduce after
the non-pipeline agg tree has been fully reduced.
By separating pipeline reduction into their own set of methods,
aggregations are free to use the final reduction for whatever
purpose without worrying about generating pipeline results
which are non-reducible
Adds `GET /_script_language` to support Kibana dynamic scripting
language selection.
Response contains whether `inline` and/or `stored` scripts are
enabled as determined by the `script.allowed_types` settings.
For each scripting language registered, such as `painless`,
`expression`, `mustache` or custom, available contexts for the language
are included as determined by the `script.allowed_contexts` setting.
Response format:
```
{
"types_allowed": [
"inline",
"stored"
],
"language_contexts": [
{
"language": "expression",
"contexts": [
"aggregation_selector",
"aggs"
...
]
},
{
"language": "painless",
"contexts": [
"aggregation_selector",
"aggs",
"aggs_combine",
...
]
}
...
]
}
```
Fixes: #49463
**Backport**
* Copying the request is not necessary here. We can simply release it once the response has been generated and a lot of `Unpooled` allocations that way
* Relates #32228
* I think the issue that preventet that PR that PR from being merged was solved by #39634 that moved the bulk index marker search to ByteBuf bulk access so the composite buffer shouldn't require many additional bounds checks (I'd argue the bounds checks we add, we save when copying the composite buffer)
* I couldn't neccessarily reproduce much of a speedup from this change, but I could reproduce a very measureable reduction in GC time with e.g. Rally's PMC (4g heap node and bulk requests of size 5k saw a reduction in young GC time by ~10% for me)
This commit fixes a number of issues with data replication:
- Local and global checkpoints are not updated after the new operations have been fsynced, but
might capture a state before the fsync. The reason why this probably went undetected for so
long is that AsyncIOProcessor is synchronous if you index one item at a time, and hence working
as intended unless you have a high enough level of concurrent indexing. As we rely in other
places on the assumption that we have an up-to-date local checkpoint in case of synchronous
translog durability, there's a risk for the local and global checkpoints not to be up-to-date after
replication completes, and that this won't be corrected by the periodic global checkpoint sync.
- AsyncIOProcessor also has another "bad" side effect here: if you index one bulk at a time, the
bulk is always first fsynced on the primary before being sent to the replica. Further, if one thread
is tasked by AsyncIOProcessor to drain the processing queue and fsync, other threads can
easily pile more bulk requests on top of that thread. Things are not very fair here, and the thread
might continue doing a lot more fsyncs before returning (as the other threads pile more and
more on top), which blocks it from returning as a replication request (e.g. if this thread is on the
primary, it blocks the replication requests to the replicas from going out, and delaying
checkpoint advancement).
This commit fixes all these issues, and also simplifies the code that coordinates all the after
write actions.
This rewrites long sort as a `DistanceFeatureQuery`, which can
efficiently skip non-competitive blocks and segments of documents.
Depending on the dataset, the speedups can be 2 - 10 times.
The optimization can be disabled with setting the system property
`es.search.rewrite_sort` to `false`.
Optimization is skipped when an index has 50% or more data with
the same value.
Optimization is done through:
1. Rewriting sort as `DistanceFeatureQuery` which can
efficiently skip non-competitive blocks and segments of documents.
2. Sorting segments according to the primary numeric sort field(#44021)
This allows to skip non-competitive segments.
3. Using collector manager.
When we optimize sort, we sort segments by their min/max value.
As a collector expects to have segments in order,
we can not use a single collector for sorted segments.
We use collectorManager, where for every segment a dedicated collector
will be created.
4. Using Lucene's shared TopFieldCollector manager
This collector manager is able to exchange minimum competitive
score between collectors, which allows us to efficiently skip
the whole segments that don't contain competitive scores.
5. When index is force merged to a single segment, #48533 interleaving
old and new segments allows for this optimization as well,
as blocks with non-competitive docs can be skipped.
Backport for #48804
Co-authored-by: Jim Ferenczi <jim.ferenczi@elastic.co>
* Make BlobStoreRepository Aware of ClusterState (#49639)
This is a preliminary to #49060.
It does not introduce any substantial behavior change to how the blob store repository
operates. What it does is to add all the infrastructure changes around passing the cluster service to the blob store, associated test changes and a best effort approach to tracking the latest repository generation on all nodes from cluster state updates. This brings a slight improvement to the consistency
by which non-master nodes (or master directly after a failover) will be able to determine the latest repository generation. It does not however do any tricky checks for the situation after a repository operation
(create, delete or cleanup) that could theoretically be used to get even greater accuracy to keep this change simple.
This change does not in any way alter the behavior of the blobstore repository other than adding a better "guess" for the value of the latest repo generation and is mainly intended to isolate the actual logical change to how the
repository operates in #49060
This commit adds a function in NodeClient that allows to track the progress
of a search request locally. Progress is tracked through a SearchProgressListener
that exposes query and fetch responses as well as partial and final reduces.
This new method can be used by modules/plugins inside a node in order to track the
progress of a local search request.
Relates #49091
This change adds a dynamic cluster setting named `indices.id_field_data.enabled`.
When set to `false` any attempt to load the fielddata for the `_id` field will fail
with an exception. The default value in this change is set to `false` in order to prevent
fielddata usage on this field for future versions but it will be set to `true` when backporting
to 7x. When the setting is set to true (manually or by default in 7x) the loading will also issue
a deprecation warning since we want to disallow fielddata entirely when https://github.com/elastic/elasticsearch/issues/26472
is implemented.
Closes#43599
All the implementations of `EsBlobStoreTestCase` use the exact same
bootstrap code that is also used by their implementation of
`EsBlobStoreContainerTestCase`.
This means all tests might as well live under `EsBlobStoreContainerTestCase`
saving a lot of code duplication. Also, there was no HDFS implementation for
`EsBlobStoreTestCase` which is now automatically resolved by moving the tests over
since there is a HDFS implementation for the container tests.
Same as #49518 pretty much but for GCS.
Fixing a few more spots where input stream can get closed
without being fully drained and adding assertions to make sure
it's always drained.
Moved the no-close stream wrapper to production code utilities since
there's a number of spots in production code where it's also useful
(will reuse it there in a follow-up).
We do not guarantee that EngineTestCase#getDocIds is called after the
engine has been externally refreshed. Hence, we trip an assertion
assertSearcherIsWarmedUp.
CI: https://gradle-enterprise.elastic.co/s/pm2at5qmfm2iu
Relates #48605
This commit ensures that even for requests that are known to be empty body
we at least attempt to read one bytes from the request body input stream.
This is done to work around the behavior in `sun.net.httpserver.ServerImpl.Dispatcher#handleEvent`
that will close a TCP/HTTP connection that does not have the `eof` flag (see `sun.net.httpserver.LeftOverInputStream#isEOF`)
set on its input stream. As far as I can tell the only way to set this flag is to do a read when there's no more bytes buffered.
This fixes the numerous connection closing issues because the `ServerImpl` stops closing connections that it thinks
weren't fully drained.
Also, I removed a now redundant drain loop in the Azure handler as well as removed the connection closing in the error handler's
drain action (this shouldn't have an effect but makes things more predictable/easier to reason about IMO).
I would suggest merging this and closing related issue after verifying that this fixes things on CI.
The way to locally reproduce the issues we're seeing in tests is to make the retry timings more aggressive in e.g. the azure tests
and move them to single digit values. This makes the retries happen quickly enough that they run into the async connecting closing
of allegedly non-eof connections by `ServerImpl` and produces the exact kinds of failures we're seeing currently.
Relates #49401, #49429
If some replica is performing a file-based recovery, then the check
assertNoSnapshottedIndexCommit would fail. We should increase the
timeout for this check so that we can wait until all recoveries done
or aborted.
Closes#49403
Fixing a few small issues found in this code:
1. We weren't reading the request headers but the response headers when checking for blob existence in the mocked single upload path
2. Error code can never be `null` removed the dead code that resulted
3. In the logging wrapper we weren't checking for `Throwable` so any failing assertions in the http mock would not show up since they
run on a thread managed by the mock http server
Strengthens the validateClusterFormed check that is used by the test infrastructure to make
sure that nodes are properly connected and know about each other. Is used in situations where
the cluster is scaled up and down, and where there previously was a network disruption that has
been healed.
Closes#49243
While we log exception in the handler, we may still miss exceptions
hgiher up the execution chain. This adds logging of exceptions to all
operations on the IO loop including connection establishment.
Relates #49401
This commit fixes the server side logic of "List Objects" operations
of Azure and S3 fixtures. Until today, the fixtures were returning a "
flat" view of stored objects and were not correctly handling the
delimiter parameter. This causes some objects listing to be wrongly
interpreted by the snapshot deletion logic in Elasticsearch which
relies on the ability to list child containers of BlobContainer (#42653)
to correctly delete stale indices.
As a consequence, the blobs were not correctly deleted from the
emulated storage service and stayed in heap until they got garbage
collected, causing CI failures like #48978.
This commit fixes the server side logic of Azure and S3 fixture when
listing objects so that it now return correct common blob prefixes as
expected by the snapshot deletion process. It also adds an after-test
check to ensure that tests leave the repository empty (besides the
root index files).
Closes#48978
This commit changes the ThreadContext to just use a regular ThreadLocal
over the lucene CloseableThreadLocal. The CloseableThreadLocal solves
issues with ThreadLocals that are no longer needed during runtime but
in the case of the ThreadContext, we need it for the runtime of the
node and it is typically not closed until the node closes, so we miss
out on the benefits that this class provides.
Additionally by removing the close logic, we simplify code in other
places that deal with exceptions and tracking to see if it happens when
the node is closing.
Closes#42577
This API call in most implementations is fairly IO heavy and slow
so it is more natural to be async in the first place.
Concretely though, this change is a prerequisite of #49060 since
determining the repository generation from the cluster state
introduces situations where this call would have to wait for other
operations to finish. Doing so in a blocking manner would break
`SnapshotResiliencyTests` and waste a thread.
Also, this sets up the possibility to in the future make use of async IO
where provided by the underlying Repository implementation.
In a follow-up `SnapshotsService#getRepositoryData` will be made async
as well (did not do it here, since it's another huge change to do so).
Note: This change for now does not alter the threading behaviour in any way (since `Repository#getRepositoryData` isn't forking) and is purely mechanical.
Make queries on the “_index” field fast-fail if the target shard is an index that doesn’t match the query expression. Part of the “canMatch” phase optimisations.
Closes#48473
ESIntegTestCase has logic to clean up static fields in a method
annotated with `@AfterClass` so that these fields do not trigger the
StaticFieldsInvariantRule. However, during the exceptional close of the
test cluster, this cleanup can be missed. The StaticFieldsInvariantRule
always runs and will attempt to inspect the size of the static fields
that were not cleaned up. If the `currentCluster` field of
ESIntegTestCase references an InternalTestCluster, this could hold a
reference to an implementation of a `Path` that comes from the
`sun.nio.fs` package, which the security manager will deny access to.
This casues additional noise to be generated since the
AccessControlException will cause the StaticFieldsInvariantRule to fail
and also be reported along with the actual exception that occurred.
This change clears the static fields of ESIntegTestCase in a finally
block inside the `@AfterClass` method to prevent this unnecessary noise.
Closes#41526
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.
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
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.
This commit changes the ESMockAPIBasedRepositoryIntegTestCase so
that HttpHandler are now wrapped in order to log any exceptions that
could be thrown when executing the server side logic in repository
integration tests.
This commits sends the cluster name and discovery naode in the transport
level handshake response. This will allow us to stop sending the
transport service level handshake request in the 8.0-8.x release cycle.
It is necessary to start sending this in 7.x so that 8.0 is guaranteed
to be communicating with a version that sends the required information.
The realtime GET API currently has erratic performance in case where a document is accessed
that has just been indexed but not refreshed yet, as the implementation will currently force an
internal refresh in that case. Refreshing can be an expensive operation, and also will block the
thread that executes the GET operation, blocking other GETs to be processed. In case of
frequent access of recently indexed documents, this can lead to a refresh storm and terrible
GET performance.
While older versions of Elasticsearch (2.x and older) did not trigger refreshes and instead opted
to read from the translog in case of realtime GET API or update API, this was removed in 5.0
(#20102) to avoid inconsistencies between values that were returned from the translog and
those returned by the index. This was partially reverted in 6.3 (#29264) to allow _update and
upsert to read from the translog again as it was easier to guarantee consistency for these, and
also brought back more predictable performance characteristics of this API. Calls to the realtime
GET API, however, would still always do a refresh if necessary to return consistent results. This
means that users that were calling realtime GET APIs to coordinate updates on client side
(realtime GET + CAS for conditional index of updated doc) would still see very erratic
performance.
This PR (together with #48707) resolves the inconsistencies between reading from translog and
index. In particular it fixes the inconsistencies that happen when requesting stored fields, which
were not available when reading from translog. In case where stored fields are requested, this
PR will reparse the _source from the translog and derive the stored fields to be returned. With
this, it changes the realtime GET API to allow reading from the translog again, avoid refresh
storms and blocking the GET threadpool, and provide overall much better and predictable
performance for this API.
We should not open new engines if a shard is closed. We break this
assumption in #45263 where we stop verifying the shard state before
creating an engine but only before swapping the engine reference.
We can fail to snapshot the store metadata or checkIndex a closed shard
if there's some IndexWriter holding the index lock.
Closes#47060
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
In repository integration tests, we drain the HTTP request body before
returning a response. Before this change this operation was done using
Streams.readFully() which uses a 8kb buffer to read the input stream, it
now uses a 1kb for the same operation. This should reduce the allocations
made during the tests and speed them up a bit on CI.
Co-authored-by: Armin Braun <me@obrown.io>
Backport of #48452.
The SAML tests have large XML documents within which various parameters
are replaced. At present, if these test are auto-formatted, the XML
documents get strung out over many, many lines, and are basically
illegible.
Fix this by using named placeholders for variables, and indent the
multiline XML documents.
The tests in `SamlSpMetadataBuilderTests` deserve a special mention,
because they include a number of certificates in Base64. I extracted
these into variables, for additional legibility.
* Extract remote "sniffing" to connection strategy (#47253)
Currently the connection strategy used by the remote cluster service is
implemented as a multi-step sniffing process in the
RemoteClusterConnection. We intend to introduce a new connection strategy
that will operate in a different manner. This commit extracts the
sniffing logic to a dedicated strategy class. Additionally, it implements
dedicated tests for this class.
Additionally, in previous commits we moved away from a world where the
remote cluster connection was mutable. Instead, when setting updates are
made, the connection is torn down and rebuilt. We still had methods and
tests hanging around for the mutable behavior. This commit removes those.
* Introduce simple remote connection strategy (#47480)
This commit introduces a simple remote connection strategy which will
open remote connections to a configurable list of user supplied
addresses. These addresses can be remote Elasticsearch nodes or
intermediate proxies. We will perform normal clustername and version
validation, but otherwise rely on the remote cluster to route requests
to the appropriate remote node.
* Make remote setting updates support diff strategies (#47891)
Currently the entire remote cluster settings infrastructure is designed
around the sniff strategy. As we introduce an additional conneciton
strategy this infrastructure needs to be modified to support it. This
commit modifies the code so that the strategy implementations will tell
the service if the connection needs to be torn down and rebuilt.
As part of this commit, we will wait 10 seconds for new clusters to
connect when they are added through the "update" settings
infrastructure.
* Make remote setting updates support diff strategies (#47891)
Currently the entire remote cluster settings infrastructure is designed
around the sniff strategy. As we introduce an additional conneciton
strategy this infrastructure needs to be modified to support it. This
commit modifies the code so that the strategy implementations will tell
the service if the connection needs to be torn down and rebuilt.
As part of this commit, we will wait 10 seconds for new clusters to
connect when they are added through the "update" settings
infrastructure.
BytesReference is currently an abstract class which is extended by
various implementations. This makes it very difficult to use the
delegation pattern. The implication of this is that our releasable
BytesReference is a PagedBytesReference type and cannot be used as a
generic releasable bytes reference that delegates to any reference type.
This commit makes BytesReference an interface and introduces an
AbstractBytesReference for common functionality.
Brings handling of out of bounds points in linestrings in line with
points. Now points with latitude above 90 and below -90 are handled
the same way as for points by adjusting the longitude by moving it by
180 degrees.
Relates to #43916
This change adds a new field `"shards"` to `RepositoryData` that contains a mapping of `IndexId` to a `String[]`. This string array can be accessed by shard id to get the generation of a shard's shard folder (i.e. the `N` in the name of the currently valid `/indices/${indexId}/${shardId}/index-${N}` for the shard in question).
This allows for creating a new snapshot in the shard without doing any LIST operations on the shard's folder. In the case of AWS S3, this saves about 1/3 of the cost for updating an empty shard (see #45736) and removes one out of two remaining potential issues with eventually consistent blob stores (see #38941 ... now only the root `index-${N}` is determined by listing).
Also and equally if not more important, a number of possible failure modes on eventually consistent blob stores like AWS S3 are eliminated by moving all delete operations to the `master` node and moving from incremental naming of shard level index-N to uuid suffixes for these blobs.
This change moves the deleting of the previous shard level `index-${uuid}` blob to the master node instead of the data node allowing for a safe and consistent update of the shard's generation in the `RepositoryData` by first updating `RepositoryData` and then deleting the now unreferenced `index-${newUUID}` blob.
__No deletes are executed on the data nodes at all for any operation with this change.__
Note also: Previous issues with hanging data nodes interfering with master nodes are completely impossible, even on S3 (see next section for details).
This change changes the naming of the shard level `index-${N}` blobs to a uuid suffix `index-${UUID}`. The reason for this is the fact that writing a new shard-level `index-` generation blob is not atomic anymore in its effect. Not only does the blob have to be written to have an effect, it must also be referenced by the root level `index-N` (`RepositoryData`) to become an effective part of the snapshot repository.
This leads to a problem if we were to use incrementing names like we did before. If a blob `index-${N+1}` is written but due to the node/network/cluster/... crashes the root level `RepositoryData` has not been updated then a future operation will determine the shard's generation to be `N` and try to write a new `index-${N+1}` to the already existing path. Updates like that are problematic on S3 for consistency reasons, but also create numerous issues when thinking about stuck data nodes.
Previously stuck data nodes that were tasked to write `index-${N+1}` but got stuck and tried to do so after some other node had already written `index-${N+1}` were prevented form doing so (except for on S3) by us not allowing overwrites for that blob and thus no corruption could occur.
Were we to continue using incrementing names, we could not do this. The stuck node scenario would either allow for overwriting the `N+1` generation or force us to continue using a `LIST` operation to figure out the next `N` (which would make this change pointless).
With uuid naming and moving all deletes to `master` this becomes a non-issue. Data nodes write updated shard generation `index-${uuid}` and `master` makes those `index-${uuid}` part of the `RepositoryData` that it deems correct and cleans up all those `index-` that are unused.
Co-authored-by: Yannick Welsch <yannick@welsch.lu>
Co-authored-by: Tanguy Leroux <tlrx.dev@gmail.com>
The code here was needlessly complicated when it
enqueued all file uploads up-front. Instead, we can
go with a cleaner worker + queue pattern here by taking
the max-parallelism from the threadpool info.
Also, I slightly simplified the rethrow and
listener (step listener is pointless when you add the callback in the next line)
handling it since I noticed that we were needlessly rethrowing in the same
code and that wasn't worth a separate PR.
Yannick Welsch
Armin Braun
Nhat Nguyen
Przemyslaw Gomulka
Stuart Tettemer
Zachary Tong
Mayya Sharipova
Jim Ferenczi
Tanguy Leroux
Jay Modi
markharwood
Rory Hunter
Henning Andersen
Tim Brooks
Igor Motov