In case if one Distributed has multiple shards, and underlying
Distributed has only one, there can be the case when the query will be
tried to process from Complete to WithMergeableStateAfterAggregation,
which is obviously wrong.
Before the following queries was running LimitBy/Distinct step on the
initator:
select distinct sharding_key from dist order by k
While this can be omitted.
Before this patch it wasn't possible to optimize simple SELECT * FROM
dist ORDER BY (w/o GROUP BY and DISTINCT) to more optimal stage
(QueryProcessingStage::WithMergeableStateAfterAggregationAndLimit),
since that code was under
allow_nondeterministic_optimize_skip_unused_shards, rework it and make
it possible.
Also now distributed_push_down_limit is respected for
optimize_distributed_group_by_sharding_key.
Next step will be to enable distributed_push_down_limit by default.
v2: fix detection of aggregates
Before we incorrectly check that columns from GROUP BY was a subset of
columns from sharding key, while this is not right, consider the
following example:
select k1, any(k2), sum(v) from remote('127.{1,2}', view(select 1 k1, 2 k2, 3 v), cityHash64(k1, k2)) group by k1
Here the columns from GROUP BY is a subset of columns from sharding key,
but the optimization cannot be applied, since there is no guarantee that
particular shard contains distinct values of k1.
So instead we should check that GROUP BY contains all columns that is
required for calculating sharding key expression, i.e.:
select k1, k2, sum(v) from remote('127.{1,2}', view(select 1 k1, 2 k2, 3 v), cityHash64(k1, k2)) group by k1, k2
- create directory monitors in parallel (this also includes rmdir in
case of directory is empty, since even if the directory is empty it
may take some time to remove it, due to waiting for journal or if the
directory is large, i.e. it had lots of files before, since remember
ext4 does not truncate the directory size on each unlink [1])
- initialize increment in parallel too (since it does readdir())
[1]: https://lore.kernel.org/linux-ext4/930A5754-5CE6-4567-8CF0-62447C97825C@dilger.ca/
This way the remote nodes will not need to send all the rows, so this
will decrease network io and also this will make queries w/
optimize_aggregation_in_order=1/LIMIT X and w/o ORDER BY faster since it
initiator will not need to read all the rows, only first X (but note
that for this you need to your data to be sharded correctly or you may
get inaccurate results).
Note, that having lots of processing stages will increase the complexity
of interpreter (it is already not that clean and simple right now).
Although using separate QueryProcessingStage looks pretty natural.
Another option is to make WithMergeableStateAfterAggregation always, but
in this case you will not be able to disable only this optimization,
i.e. if there will be some issue with it.
v2: fix OFFSET
v3: convert 01814_distributed_push_down_limit test to .sh and add retries
v4: add test with OFFSET
v5: add new query stage into the bash completion
v6/tests: use LIMIT O,L syntax over LIMIT L OFFSET O since it is broken in ANTLR parser
https://clickhouse-test-reports.s3.yandex.net/23027/a18a06399b7aeacba7c50b5d1e981ada5df19745/functional_stateless_tests_(antlr_debug).html#fail1
v7/tests: set use_hedged_requests to 0, to avoid excessive log entries on retries
https://clickhouse-test-reports.s3.yandex.net/23027/a18a06399b7aeacba7c50b5d1e981ada5df19745/functional_stateless_tests_flaky_check_(address).html#fail1
v2: fix optimize_skip_unused_shards_rewrite_in for sharding_key wrapped into function
v3: fix column name for optimize_skip_unused_shards_rewrite_in
v4: fix optimize_skip_unused_shards_rewrite_in with Null
v5:
- squash with Remove query argument for IStreamFactory::createForShard()
- use proper column after function execution (using sharding_key_column_name)
- update the test reference since (X) now is tuple(X)
Add two new settings for the Distributed engine:
- bytes_to_delay_insert
- max_delay_to_insert
If at the beginning of INSERT there will be too much pending data, more
then bytes_to_delay_insert, then the INSERT will wait until it will be
shrinked, and not more then max_delay_to_insert seconds.
If after this there will be still too much pending, it will throw an
exception.
Also new profile events were added (by analogy to the MergeTree):
- DistributedDelayedInserts (although you can use system.errors instead
of this, but still)
- DistributedRejectedInserts
- DistributedDelayedInsertsMilliseconds
So now system.distribution_queue will show accurate statistics, so tests
does not requires sleep anymore.
But note that with too much distributed pending this will iterate over
all directories.
Right now with distributed_directory_monitor_batch_inserts=1 and
insert_distributed_sync=0 INSERT into Distributed table will store
blocks that should be sent to remote (and in case of
prefer_localhost_replica=0 to the localhost too) on the local
filesystem, and sent it in background.
However there is no limit for this storage, and if the remote is
unavailable (or some other error), these pending blocks may take
significant space, and this is not always desired behaviour.
Add new Distributed setting - bytes_to_throw_insert, that will set the
limit for how much pending bytes is allowed, if the limit will be
reached an exception will be throw.
By default was set to 0, to avoid surprises.
Right now SYSTEM FLUSH DISTRIBUTED will block:
- INSERT into this Distributed table (requireDirectoryMonitor())
- SELECT * FROM system.distribution_queue
* add the query data deduplication excluding duplicated parts in MergeTree family engines.
query deduplication is based on parts' UUID which should be enabled first with merge_tree setting
assign_part_uuids=1
allow_experimental_query_deduplication setting is to enable part deduplication, default ot false.
data part UUID is a mechanism of giving a data part a unique identifier.
Having UUID and deduplication mechanism provides a potential of moving parts
between shards preserving data consistency on a read path:
duplicated UUIDs will cause root executor to retry query against on of the replica explicitly
asking to exclude encountered duplicated fingerprints during a distributed query execution.
NOTE: this implementation don't provide any knobs to lock part and hence its UUID. Any mutations/merge will
update part's UUID.
* add _part_uuid virtual column, allowing to use UUIDs in predicates.
Signed-off-by: Aleksei Semiglazov <asemiglazov@cloudflare.com>
address comments
Two new settings (by analogy with MergeTree family) has been added:
- `fsync_after_insert` - Do fsync for every inserted. Will decreases
performance of inserts.
- `fsync_tmp_directory` - Do fsync for temporary directory (that is used
for async INSERT only) after all part operations (writes, renames,
etc.).
Refs: #17380 (p1)
