With this commit, we're introducing the Adaptive Executor.
The commit message consists of two distinct sections. The first part explains
how the executor works. The second part consists of the commit messages of
the individual smaller commits that resulted in this commit. The readers
can search for the each of the smaller commit messages on
https://github.com/citusdata/citus and can learn more about the history
of the change.
/*-------------------------------------------------------------------------
*
* adaptive_executor.c
*
* The adaptive executor executes a list of tasks (queries on shards) over
* a connection pool per worker node. The results of the queries, if any,
* are written to a tuple store.
*
* The concepts in the executor are modelled in a set of structs:
*
* - DistributedExecution:
* Execution of a Task list over a set of WorkerPools.
* - WorkerPool
* Pool of WorkerSessions for the same worker which opportunistically
* executes "unassigned" tasks from a queue.
* - WorkerSession:
* Connection to a worker that is used to execute "assigned" tasks
* from a queue and may execute unasssigned tasks from the WorkerPool.
* - ShardCommandExecution:
* Execution of a Task across a list of placements.
* - TaskPlacementExecution:
* Execution of a Task on a specific placement.
* Used in the WorkerPool and WorkerSession queues.
*
* Every connection pool (WorkerPool) and every connection (WorkerSession)
* have a queue of tasks that are ready to execute (readyTaskQueue) and a
* queue/set of pending tasks that may become ready later in the execution
* (pendingTaskQueue). The tasks are wrapped in a ShardCommandExecution,
* which keeps track of the state of execution and is referenced from a
* TaskPlacementExecution, which is the data structure that is actually
* added to the queues and describes the state of the execution of a task
* on a particular worker node.
*
* When the task list is part of a bigger distributed transaction, the
* shards that are accessed or modified by the task may have already been
* accessed earlier in the transaction. We need to make sure we use the
* same connection since it may hold relevant locks or have uncommitted
* writes. In that case we "assign" the task to a connection by adding
* it to the task queue of specific connection (in
* AssignTasksToConnections). Otherwise we consider the task unassigned
* and add it to the task queue of a worker pool, which means that it
* can be executed over any connection in the pool.
*
* A task may be executed on multiple placements in case of a reference
* table or a replicated distributed table. Depending on the type of
* task, it may not be ready to be executed on a worker node immediately.
* For instance, INSERTs on a reference table are executed serially across
* placements to avoid deadlocks when concurrent INSERTs take conflicting
* locks. At the beginning, only the "first" placement is ready to execute
* and therefore added to the readyTaskQueue in the pool or connection.
* The remaining placements are added to the pendingTaskQueue. Once
* execution on the first placement is done the second placement moves
* from pendingTaskQueue to readyTaskQueue. The same approach is used to
* fail over read-only tasks to another placement.
*
* Once all the tasks are added to a queue, the main loop in
* RunDistributedExecution repeatedly does the following:
*
* For each pool:
* - ManageWorkPool evaluates whether to open additional connections
* based on the number unassigned tasks that are ready to execute
* and the targetPoolSize of the execution.
*
* Poll all connections:
* - We use a WaitEventSet that contains all (non-failed) connections
* and is rebuilt whenever the set of active connections or any of
* their wait flags change.
*
* We almost always check for WL_SOCKET_READABLE because a session
* can emit notices at any time during execution, but it will only
* wake up WaitEventSetWait when there are actual bytes to read.
*
* We check for WL_SOCKET_WRITEABLE just after sending bytes in case
* there is not enough space in the TCP buffer. Since a socket is
* almost always writable we also use WL_SOCKET_WRITEABLE as a
* mechanism to wake up WaitEventSetWait for non-I/O events, e.g.
* when a task moves from pending to ready.
*
* For each connection that is ready:
* - ConnectionStateMachine handles connection establishment and failure
* as well as command execution via TransactionStateMachine.
*
* When a connection is ready to execute a new task, it first checks its
* own readyTaskQueue and otherwise takes a task from the worker pool's
* readyTaskQueue (on a first-come-first-serve basis).
*
* In cases where the tasks finish quickly (e.g. <1ms), a single
* connection will often be sufficient to finish all tasks. It is
* therefore not necessary that all connections are established
* successfully or open a transaction (which may be blocked by an
* intermediate pgbouncer in transaction pooling mode). It is therefore
* essential that we take a task from the queue only after opening a
* transaction block.
*
* When a command on a worker finishes or the connection is lost, we call
* PlacementExecutionDone, which then updates the state of the task
* based on whether we need to run it on other placements. When a
* connection fails or all connections to a worker fail, we also call
* PlacementExecutionDone for all queued tasks to try the next placement
* and, if necessary, mark shard placements as inactive. If a task fails
* to execute on all placements, the execution fails and the distributed
* transaction rolls back.
*
* For multi-row INSERTs, tasks are executed sequentially by
* SequentialRunDistributedExecution instead of in parallel, which allows
* a high degree of concurrency without high risk of deadlocks.
* Conversely, multi-row UPDATE/DELETE/DDL commands take aggressive locks
* which forbids concurrency, but allows parallelism without high risk
* of deadlocks. Note that this is unrelated to SEQUENTIAL_CONNECTION,
* which indicates that we should use at most one connection per node, but
* can run tasks in parallel across nodes. This is used when there are
* writes to a reference table that has foreign keys from a distributed
* table.
*
* Execution finishes when all tasks are done, the query errors out, or
* the user cancels the query.
*
*-------------------------------------------------------------------------
*/
All the commits involved here:
* Initial unified executor prototype
* Latest changes
* Fix rebase conflicts to master branch
* Add missing variable for assertion
* Ensure that master_modify_multiple_shards() returns the affectedTupleCount
* Adjust intermediate result sizes
The real-time executor uses COPY command to get the results
from the worker nodes. Unified executor avoids that which
results in less data transfer. Simply adjust the tests to lower
sizes.
* Force one connection per placement (or co-located placements) when requested
The existing executors (real-time and router) always open 1 connection per
placement when parallel execution is requested.
That might be useful under certain circumstances:
(a) User wants to utilize as much as CPUs on the workers per
distributed query
(b) User has a transaction block which involves COPY command
Also, lots of regression tests rely on this execution semantics.
So, we'd enable few of the tests with this change as well.
* For parameters to be resolved before using them
For the details, see PostgreSQL's copyParamList()
* Unified executor sorts the returning output
* Ensure that unified executor doesn't ignore sequential execution of DDLJob's
Certain DDL commands, mainly creating foreign keys to reference tables,
should be executed sequentially. Otherwise, we'd end up with a self
distributed deadlock.
To overcome this situaiton, we set a flag `DDLJob->executeSequentially`
and execute it sequentially. Note that we have to do this because
the command might not be called within a transaction block, and
we cannot call `SetLocalMultiShardModifyModeToSequential()`.
This fixes at least two test: multi_insert_select_on_conflit.sql and
multi_foreign_key.sql
Also, I wouldn't mind scattering local `targetPoolSize` variables within
the code. The reason is that we'll soon have a GUC (or a global
variable based on a GUC) that'd set the pool size. In that case, we'd
simply replace `targetPoolSize` with the global variables.
* Fix 2PC conditions for DDL tasks
* Improve closing connections that are not fully established in unified execution
* Support foreign keys to reference tables in unified executor
The idea for supporting foreign keys to reference tables is simple:
Keep track of the relation accesses within a transaction block.
- If a parallel access happens on a distributed table which
has a foreign key to a reference table, one cannot modify
the reference table in the same transaction. Otherwise,
we're very likely to end-up with a self-distributed deadlock.
- If an access to a reference table happens, and then a parallel
access to a distributed table (which has a fkey to the reference
table) happens, we switch to sequential mode.
Unified executor misses the function calls that marks the relation
accesses during the execution. Thus, simply add the necessary calls
and let the logic kick in.
* Make sure to close the failed connections after the execution
* Improve comments
* Fix savepoints in unified executor.
* Rebuild the WaitEventSet only when necessary
* Unclaim connections on all errors.
* Improve failure handling for unified executor
- Implement the notion of errorOnAnyFailure. This is similar to
Critical Connections that the connection managament APIs provide
- If the nodes inside a modifying transaction expand, activate 2PC
- Fix few bugs related to wait event sets
- Mark placement INACTIVE during the execution as much as possible
as opposed to we do in the COMMIT handler
- Fix few bugs related to scheduling next placement executions
- Improve decision on when to use 2PC
Improve the logic to start a transaction block for distributed transactions
- Make sure that only reference table modifications are always
executed with distributed transactions
- Make sure that stored procedures and functions are executed
with distributed transactions
* Move waitEventSet to DistributedExecution
This could also be local to RunDistributedExecution(), but in that case
we had to mark it as "volatile" to avoid PG_TRY()/PG_CATCH() issues, and
cast it to non-volatile when doing WaitEventSetFree(). We thought that
would make code a bit harder to read than making this non-local, so we
move it here. See comments for PG_TRY() in postgres/src/include/elog.h
and "man 3 siglongjmp" for more context.
* Fix multi_insert_select test outputs
Two things:
1) One complex transaction block is now supported. Simply update
the test output
2) Due to dynamic nature of the unified executor, the orders of
the errors coming from the shards might change (e.g., all of
the queries on the shards would fail, but which one appears
on the error message?). To fix that, we simply added it to
our shardId normalization tool which happens just before diff.
* Fix subeury_and_cte test
The error message is updated from:
failed to execute task
To:
more than one row returned by a subquery or an expression
which is a lot clearer to the user.
* Fix intermediate_results test outputs
Simply update the error message from:
could not receive query results
to
result "squares" does not exist
which makes a lot more sense.
* Fix multi_function_in_join test
The error messages update from:
Failed to execute task XXX
To:
function f(..) does not exist
* Fix multi_query_directory_cleanup test
The unified executor does not create any intermediate files.
* Fix with_transactions test
A test case that just started to work fine
* Fix multi_router_planner test outputs
The error message is update from:
Could not receive query results
To:
Relation does not exists
which is a lot more clearer for the users
* Fix multi_router_planner_fast_path test
The error message is update from:
Could not receive query results
To:
Relation does not exists
which is a lot more clearer for the users
* Fix isolation_copy_placement_vs_modification by disabling select_opens_transaction_block
* Fix ordering in isolation_multi_shard_modify_vs_all
* Add executor locks to unified executor
* Make sure to allocate enought WaitEvents
The previous code was missing the waitEvents for the latch and
postmaster death.
* Fix rebase conflicts for master rebase
* Make sure that TRUNCATE relies on unified executor
* Implement true sequential execution for multi-row INSERTS
Execute the individual tasks executed one by one. Note that this is different than
MultiShardConnectionType == SEQUENTIAL_CONNECTION case (e.g., sequential execution
mode). In that case, running the tasks across the nodes in parallel is acceptable
and implemented in that way.
However, the executions that are qualified here would perform poorly if the
tasks across the workers are executed in parallel. We currently qualify only
one class of distributed queries here, multi-row INSERTs. If we do not enforce
true sequential execution, concurrent multi-row upserts could easily form
a distributed deadlock when the upserts touch the same rows.
* Remove SESSION_LIFESPAN flag in unified_executor
* Apply failure test updates
We've changed the failure behaviour a bit, and also the error messages
that show up to the user. This PR covers majority of the updates.
* Unified executor honors citus.node_connection_timeout
With this commit, unified executor errors out if even
a single connection cannot be established within
citus.node_connection_timeout.
And, as a side effect this fixes failure_connection_establishment
test.
* Properly increment/decrement pool size variables
Before this commit, the idle and active connection
counts were not properly calculated.
* insert_select_executor goes through unified executor.
* Add missing file for task tracker
* Modify ExecuteTaskListExtended()'s signature
* Sort output of INSERT ... SELECT ... RETURNING
* Take partition locks correctly in unified executor
* Alternative implementation for force_max_query_parallelization
* Fix compile warnings in unified executor
* Fix style issues
* Decrement idleConnectionCount when idle connection is lost
* Always rebuild the wait event sets
In the previous implementation, on waitFlag changes, we were only
modifying the wait events. However, we've realized that it might
be an over optimization since (a) we couldn't see any performance
benefits (b) we see some errors on failures and because of (a)
we prefer to disable it now.
* Make sure to allocate enough sized waitEventSet
With multi-row INSERTs, we might have more sessions than
task*workerCount after few calls of RunDistributedExecution()
because the previous sessions would also be alive.
Instead, re-allocate events when the connectino set changes.
* Implement SELECT FOR UPDATE on reference tables
On master branch, we do two extra things on SELECT FOR UPDATE
queries on reference tables:
- Acquire executor locks
- Execute the query on all replicas
With this commit, we're implementing the same logic on the
new executor.
* SELECT FOR UPDATE opens transaction block even if SelectOpensTransactionBlock disabled
Otherwise, users would be very confused and their logic is very likely
to break.
* Fix build error
* Fix the newConnectionCount calculation in ManageWorkerPool
* Fix rebase conflicts
* Fix minor test output differences
* Fix citus indent
* Remove duplicate sorts that is added with rebase
* Create distributed table via executor
* Fix wait flags in CheckConnectionReady
* failure_savepoints output for unified executor.
* failure_vacuum output (pg 10) for unified executor.
* Fix WaitEventSetWait timeout in unified executor
* Stabilize failure_truncate test output
* Add an ORDER BY to multi_upsert
* Fix regression test outputs after rebase to master
* Add executor.c comment
* Rename executor.c to adaptive_executor.c
* Do not schedule tasks if the failed placement is not ready to execute
Before the commit, we were blindly scheduling the next placement executions
even if the failed placement is not on the ready queue. Now, we're ensuring
that if failed placement execution is on a failed pool or session where the
execution is on the pendingQueue, we do not schedule the next task. Because
the other placement execution should be already running.
* Implement a proper custom scan node for adaptive executor
- Switch between the executors, add GUC to set the pool size
- Add non-adaptive regression test suites
- Enable CIRCLE CI for non-adaptive tests
- Adjust test output files
* Add slow start interval to the executor
* Expose max_cached_connection_per_worker to user
* Do not start slow when there are cached connections
* Consider ExecutorSlowStartInterval in NextEventTimeout
* Fix memory issues with ReceiveResults().
* Disable executor via TaskExecutorType
* Make sure to execute the tests with the other executor
* Use task_executor_type to enable-disable adaptive executor
* Remove useless code
* Adjust the regression tests
* Add slow start regression test
* Rebase to master
* Fix test failures in adaptive executor.
* Rebase to master - 2
* Improve comments & debug messages
* Set force_max_query_parallelization in isolation_citus_dist_activity
* Force max parallelization for creating shards when asked to use exclusive connection.
* Adjust the default pool size
* Expand description of max_adaptive_executor_pool_size GUC
* Update warnings in FinishRemoteTransactionCommit()
* Improve session clean up at the end of execution
Explicitly list all the states that the execution might end,
otherwise warn.
* Remove MULTI_CONNECTION_WAIT_RETRY which is not used at all
* Add more ORDER BYs to multi_mx_partitioning
In this context, we define "Fast Path Planning for SELECT" as trivial
queries where Citus can skip relying on the standard_planner() and
handle all the planning.
For router planner, standard_planner() is mostly important to generate
the necessary restriction information. Later, the restriction information
generated by the standard_planner is used to decide whether all the shards
that a distributed query touches reside on a single worker node. However,
standard_planner() does a lot of extra things such as cost estimation and
execution path generations which are completely unnecessary in the context
of distributed planning.
There are certain types of queries where Citus could skip relying on
standard_planner() to generate the restriction information. For queries
in the following format, Citus does not need any information that the
standard_planner() generates:
SELECT ... FROM single_table WHERE distribution_key = X; or
DELETE FROM single_table WHERE distribution_key = X; or
UPDATE single_table SET value_1 = value_2 + 1 WHERE distribution_key = X;
Note that the queries might not be as simple as the above such that
GROUP BY, WINDOW FUNCIONS, ORDER BY or HAVING etc. are all acceptable. The
only rule is that the query is on a single distributed (or reference) table
and there is a "distribution_key = X;" in the WHERE clause. With that, we
could use to decide the shard that a distributed query touches reside on
a worker node.
* Change worker_hash_partition_table() such that the
divergence between Citus planner's hashing and
worker_hash_partition_table() becomes the same.
* Rename single partitioning to single range partitioning.
* Add single hash repartitioning. Basically, logical planner
treats single hash and range partitioning almost equally.
Physical planner, on the other hand, treats single hash and
dual hash repartitioning almost equally (except for JoinPruning).
* Add a new GUC to enable this feature
After this change all the logic related to shard data fetch logic
will be removed. Planner won't plan any ShardFetchTask anymore.
Shard fetch related steps in real time executor and task-tracker
executor have been removed.
With #1804 (and related PRs), Citus gained the ability to
plan subqueries that are not safe to pushdown.
There are two high-level requirements for pushing down subqueries:
* Individual subqueries that require a merge step (i.e., GROUP BY
on non-distribution key, or LIMIT in the subquery etc). We've
handled such subqueries via #1876.
* Combination of subqueries that are not joined on distribution keys.
This commit aims to recursively plan some of such subqueries to make
the whole query safe to pushdown.
The main logic behind non colocated subquery joins is that we pick
an anchor range table entry and check for distribution key equality
of any other subqueries in the given query. If for a given subquery,
we cannot find distribution key equality with the anchor rte, we
recursively plan that subquery.
We also used a hacky solution for picking relations as the anchor range
table entries. The hack is that we wrap them into a subquery. This is only
necessary since some of the attribute equivalance checks are based on
queries rather than range table entries.
With this commit, Citus recursively plans subqueries that
are not safe to pushdown, in other words, requires a merge
step.
The algorithm is simple: Recursively traverse the query from bottom
up (i.e., bottom meaning the leaf queries). On each level, check
whether the query is safe to pushdown (or a single repartition
subquery). If the answer is yes, do not touch that subquery. If the
answer is no, plan the subquery seperately (i.e., create a subPlan
for it) and replace the subquery with a call to
`read_intermediate_results(planId, subPlanId)`. During the the
execution, run the subPlans first, and make them avaliable to the
next query executions.
Some of the queries hat this change allows us:
* Subqueries with LIMIT
* Subqueries with GROUP BY/DISTINCT on non-partition keys
* Subqueries involving re-partition joins, router queries
* Mixed usage of subqueries and CTEs (i.e., use CTEs in
subqueries as well). Nested subqueries as long as we
support the subquery inside the nested subquery.
* Subqueries with local tables (i.e., those subqueries
has the limitation that they have to be leaf subqueries)
* VIEWs on the distributed tables just works (i.e., the
limitations mentioned below still applies to views)
Some of the queries that is still NOT supported:
* Corrolated subqueries that are not safe to pushdown
* Window function on non-partition keys
* Recursively planned subqueries or CTEs on the outer
side of an outer join
* Only recursively planned subqueries and CTEs in the FROM
(i.e., not any distributed tables in the FROM) and subqueries
in WHERE clause
* Subquery joins that are not on the partition columns (i.e., each
subquery is individually joined on partition keys but not the upper
level subquery.)
* Any limitation that logical planner applies such as aggregate
distincts (except for count) when GROUP BY is on non-partition key,
or array_agg with ORDER BY
* Support for subqueries in WHERE clause
This commit enables subqueries in WHERE clause to be pushed down
by the subquery pushdown logic.
The support covers:
- Correlated subqueries with IN, NOT IN, EXISTS, NOT EXISTS,
operator expressions such as (>, <, =, ALL, ANY etc.)
- Non-correlated subqueries with (partition_key) IN (SELECT partition_key ..)
(partition_key) =ANY (SELECT partition_key ...)
Note that this commit heavily utilizes the attribute equivalence logic introduced
in the 1cb6a34ba8. In general, this commit mostly
adjusts the logical planner not to error out on the subqueries in WHERE clause.
* Improve error checks for subquery pushdown and INSERT ... SELECT
Since we allow subqueries in WHERE clause with the previous commit,
we should apply the same limitations to those subqueries.
With this commit, we do not iterate on each subquery one by one.
Instead, we extract all the subqueries and apply the checks directly
on those subqueries. The aim of this change is to (i) Simplify the
code (ii) Make it close to the checks on INSERT .. SELECT code base.
* Extend checks for unresolved paramaters to include SubLinks
With the presence of subqueries in where clause (i.e., SubPlans on the
query) the existing way for checking unresolved parameters fail. The
reason is that the parameters for SubPlans are kept on the parent plan not
on the query itself (see primnodes.h for the details).
With this commit, instead of checking SubPlans on the modified plans
we start to use originalQuery, where SubLinks represent the subqueries
in where clause. The unresolved parameters can be found on the SubLinks.
* Apply code-review feedback
* Remove unnecessary copying of shard interval list
This commit removes unnecessary copying of shard interval list. Note
that there are no copyObject function implemented for shard intervals.
Custom Scan is a node in the planned statement which helps external providers
to abstract data scan not just for foreign data wrappers but also for regular
relations so you can benefit your version of caching or hardware optimizations.
This sounds like only an abstraction on the data scan layer, but we can use it
as an abstraction for our distributed queries. The only thing we need to do is
to find distributable parts of the query, plan for them and replace them with
a Citus Custom Scan. Then, whenever PostgreSQL hits this custom scan node in
its Vulcano style execution, it will call our callback functions which run
distributed plan and provides tuples to the upper node as it scans a regular
relation. This means fewer code changes, fewer bugs and more supported features
for us!
First, in the distributed query planner phase, we create a Custom Scan which
wraps the distributed plan. For real-time and task-tracker executors, we add
this custom plan under the master query plan. For router executor, we directly
pass the custom plan because there is not any master query. Then, we simply let
the PostgreSQL executor run this plan. When it hits the custom scan node, we
call the related executor parts for distributed plan, fill the tuple store in
the custom scan and return results to PostgreSQL executor in Vulcano style,
a tuple per XXX_ExecScan() call.
* Modify planner to utilize Custom Scan node.
* Create different scan methods for different executors.
* Use native PostgreSQL Explain for master part of queries.
- Break CheckShardPlacements into multiple functions (The most important
is MarkFailedShardPlacements), so that we can get rid of the global
CoordinatedTransactionUses2PC.
- Call MarkFailedShardPlacements in the router executor, so we mark
shards as invalid and stop using them while inside transaction blocks.
With this change we start to error out on router planner queries where a common table
expression with data-modifying statement is present. We already do not support if
there is a data-modifying statement using result of the CTE, now we also error out
if CTE itself is data-modifying statement.
Router planner already handles cases when all shards
are pruned out. This is about missing test cases. Notice that
"column is null" and "column = null" have different shard
pruning behavior.
We used to disable router planner and executor
when task executor is set to task-tracker.
This change enables router planning and execution
at all times regardless of task execution mode.
We are introducing a hidden flag enable_router_execution
to enable/disable router execution. Its default value is
true. User may disable router planning by setting it to false.
Adds support for PostgreSQL 9.6 by copying in the requisite ruleutils
file and refactoring the out/readfuncs code to flexibly support the
old-style copy/pasted out/readfuncs (prior to 9.6) or use extensible
node APIs (in 9.6 and higher).
Most version-specific code within this change is only needed to set new
fields in the AggRef nodes we build for aggregations. Version-specific
test output files were added in certain cases, though in most they were
not necessary. Each such file begins by e.g. printing the major version
in order to clarify its purpose.
The comment atop citus_nodes.h details how to add support for new nodes
for when that becomes necessary.
PostgreSQL 9.5.4 stopped calling planner for materialized view create
command when NO DATA option is provided.
This causes our test to behave differently between pre-9.5.4 and 9.5.4.
We can now support richer set of queries in router planner.
This allow us to support CTEs, joins, window function, subqueries
if they are known to be executed at a single worker with a single
task (all tables are filtered down to a single shard and a single
worker contains all table shards referenced in the query).
Fixes : #501
Fixes#271
This change sets ShardIds and JobIds for each test case. Before this change,
when a new test that somehow increments Job or Shard IDs is added, then
the tests after the new test should be updated.
ShardID and JobID sequences are set at the beginning of each file with the
following commands:
```
ALTER SEQUENCE pg_catalog.pg_dist_shardid_seq RESTART 290000;
ALTER SEQUENCE pg_catalog.pg_dist_jobid_seq RESTART 290000;
```
ShardIds and JobIds are multiples of 10000. Exceptions are:
- multi_large_shardid: shardid and jobid sequences are set to much larger values
- multi_fdw_large_shardid: same as above
- multi_join_pruning: Causes a race condition with multi_hash_pruning since
they are run in parallel.
- non-router plannable queries can be executed
by router executor if they satisfy the criteria
- router executor is removed from configuration,
now task executor can not be set to router
- removed some tests that error out for router executor
Marco Slot
Philip Dubé
Philip Dubé
Önder Kalacı
Murat Tuncer
velioglu
Metin Doslu
Brian Cloutier
mehmet furkan şahin
Jason Petersen
Andres Freund
Brian Cloutier
Burak Yucesoy
Eren