So far citus used postgres' predicate proofing logic for shard
pruning, except for INSERT and COPY which were already optimized for
speed. That turns out to be too slow:
* Shard pruning for SELECTs is currently O(#shards), because
PruneShardList calls predicate_refuted_by() for every
shard. Obviously using an O(N) type algorithm for general pruning
isn't good.
* predicate_refuted_by() is quite expensive on its own right. That's
primarily because it's optimized for doing a single refutation
proof, rather than performing the same proof over and over.
* predicate_refuted_by() does not keep persistent state (see 2.) for
function calls, which means that a lot of syscache lookups will be
performed. That's particularly bad if the partitioning key is a
composite key, because without a persistent FunctionCallInfo
record_cmp() has to repeatedly look-up the type definition of the
composite key. That's quite expensive.
Thus replace this with custom-code that works in two phases:
1) Search restrictions for constraints that can be pruned upon
2) Use those restrictions to search for matching shards in the most
efficient manner available:
a) Binary search / Hash Lookup in case of hash partitioned tables
b) Binary search for equal clauses in case of range or append
tables without overlapping shards.
c) Binary search for inequality clauses, searching for both lower
and upper boundaries, again in case of range or append
tables without overlapping shards.
d) exhaustive search testing each ShardInterval
My measurements suggest that we are considerably, often orders of
magnitude, faster than the previous solution, even if we have to fall
back to exhaustive pruning.
This determines whether it's possible to perform binary search on
sortedShardIntervalArray or not. If e.g. two shards have overlapping
ranges, that'd be prohibitive.
That'll be useful in later commit introducing faster shard pruning.
That's useful when comparing values a hash-partitioned table is
filtered by. The existing shardIntervalCompareFunction is about
comparing hashed values, not unhashed ones.
The added btree opclass function is so we can get a comparator
back. This should be changed much more widely, but is not necessary so
far.
Previously we, unnecessarily, used a the first shard's type
information to to look up the comparison function. But that
information is already available, so use it. That's helpful because
we sometimes want to access the comparator function even if there's no
shards.
With this commit, we started to send explain queries within a savepoint. After
running explain query, we rollback to savepoint. This saves us from side effects
of EXPLAIN ANALYZE on DML queries.
All callers fetch a cache entry and extract/compute arguments for the
eventual FindShardInterval call, so it makes more sense to refactor
into that function itself; this solves the use-after-free bug, too.
Soon shard pruning will be optimized not to generally work linearly
anymore. Thus we can't print the pruned shard intervals as currently
done anymore.
The current printing of shard ids also prevents us from running tests
in parallel, as otherwise shard ids aren't linearly numbered.
Pretty straightforward. Had some concerns about locking, but due to the
fact that all distributed operations use either some level of deparsing
or need to enumerate column names, they all block during any concurrent
column renames (due to the AccessExclusive lock).
In addition, I had some misgivings about permitting renames of the dis-
tribution column, but nothing bad comes from just allowing them.
Finally, I tried to trigger any sort of error using prepared statements
and could not trigger any errors not also exhibited by plain PostgreSQL
tables.
With this change, we set to default value of isactive column to true so that
upgrading users all nodes will be marked as active to not break their environment.
With this change we add an option to add a node without replicating all reference
tables to that node. If a node is added with this option, we mark the node as
inactive and no queries will sent to that node.
We also added two new UDFs;
- master_activate_node(host, port):
- marks node as active and replicates all reference tables to that node
- master_add_inactive_node(host, port):
- only adds node to pg_dist_node
Before this commit, we were erroring out for queries containing parameterized SQL functions
like 'SELECT parameterized_sql_query(value)' as we should, however we were returning wrong
results for queries like 'SELECT * FROM parameterized_sql_query(value)'. With this commit
we started to error out on such queries too.
In this PR, we aim to deduce whether each of the RTE_RELATION
is joined with at least on another RTE_RELATION on their partition keys. If each
RTE_RELATION follows the above rule, we can conclude that all RTE_RELATIONs are
joined on their partition keys.
In order to do that, we invented a new equivalence class namely:
AttributeEquivalenceClass. In very simple words, a AttributeEquivalenceClass is
identified by an unique id and consists of a list of AttributeEquivalenceMembers.
Each AttributeEquivalenceMember is designed to identify attributes uniquely within the
whole query. The necessity of this arise since varno attributes are defined within
a single level of a query. Instead, here we want to identify each RTE_RELATION uniquely
and try to find equality among each RTE_RELATION's partition key.
Whenever we find an equality clause A = B, where both A and B originates from
relation attributes (i.e., not random expressions), we create an
AttributeEquivalenceClass to record this knowledge. If we later find another
equivalence B = C, we create another AttributeEquivalenceClass. Finally, we can
apply transitity rules and generate a new AttributeEquivalenceClass which includes
A, B and C.
Note that equality among the members are identified by the varattno and rteIdentity.
Each equality among RTE_RELATION is saved using an AttributeEquivalenceClass where
each member attribute is identified by a AttributeEquivalenceMember. In the final
step, we try generate a common attribute equivalence class that holds as much as
AttributeEquivalenceMembers whose attributes are a partition keys.
This was getting pretty long and complex in the context of the main
utility hook. Moved out the checks for what should skip Citus process-
ing and what should have version checks performed.
With this change, we start to error out if loaded citus binaries does not match
the available major version or installed citus extension version. In this case
we force user to restart the server or run ALTER EXTENSION depending on the
situation
Thought this looked slightly nicer than the default behavior.
Changed preventTransaction to concurrent to be clearer that this code
path presently affects CONCURRENTLY code only.
Coordinator code marks index as invalid as a base, set it as valid in a
transactional layer atop that base, then proceeds with worker commands.
If a worker command has problems, the rollback results in an index with
isvalid = false. If everything succeeds, the user sees a valid index.
Before this commit, in certain cases router planner allowed pushing
down JOINs that are not on the partition keys.
With @anarazel's suggestion, we change the logic to use uninstantiated
parameter. Previously, the planner was traversing on the restriction
information and once it finds the parameter, it was replacing it with
the shard range. With this commit, instead of traversing the restrict
infos, the planner explicitly checks for the equivalence of the relation
partition key with the uninstantiated parameter. If finds an equivalence,
it adds the restrictions. In this way, we have more control over the
queries that are pushed down.
Some tests relied on worker errors though local commands were invalid.
Fixed those by ensuring preconditions were met to have command work
correctly. Otherwise most test changes are related to slight changes
in local/remote error ordering.
With this commit, we add the range table list of the original query to our
custom plan. Therefore, PostgreSQL can check relations in the original query
for access permissions and error out if the proper access is not granted.
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.
Previously we'd segfault in PQisnonblocking() which, contrary to other
libpq calls, doesn't handle a NULL PQconn (because there'd be no
appropriate return value for that).
cr: @jasonmp85
Delete operation is blocked for any table distributed by hash using master_apply_delete_command. Suggested master_modify_multiple_shards command as a hint.
During later work the transaction debug output will change (as it will
in postgres 10), which makes it hard to see actual changes in the
INSERT ... SELECT ... test. Reduce to DEBUG2 after changing a debug
message to that log level.
This change ignores `citus.replication_model` setting and uses the
statement based replication in
- Tables distributed via the old `master_create_distributed_table` function
- Append and range partitioned tables, even if created via
`create_distributed_table` function
This seems like the easiest solution to #1191, without changing the existing
behavior and harming existing users with custom scripts.
This change also prevents RF>1 on streaming replicated tables on `master_create_worker_shards`
Prior to this change, `master_create_worker_shards` command was not checking
the replication model of the target table, thus allowing RF>1 with streaming
replicated tables. With this change, `master_create_worker_shards` errors
out on the case.
Add a call to RemoteTransactionBeginIfNecessary so that BEGIN is
actually sent to the remote connections. This means that ROLLBACK and
Ctrl-C are respected and don't leave the table in a partial state.
This change allows users to drop sequences on MX workers. Previously, Citus didn't allow dropping
sequences on MX workers because it could cause shards to be dropped if `DROP SEQUENCE ... CASCADE`
is used. We now allow that since allowing sequence creation but not dropping hurts user experience
and also may cause problems with custom Citus solutions.
- 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 DropShards function started to use new connection API. DropShards
function is used by DROP TABLE, master_drop_all_shards and master_apply_delete_command,
therefore all of these functions now support transactional operations. In DropShards
function, if we cannot reach a node, we mark shard state of related placements as
FILE_TO_DELETE and continue to drop remaining shards; however if any error occurs after
establishing the connection, we ROLLBACK whole operation.
Before this change, when a transaction failed, we update related placements shard states
to FILE_INACTIVE during XACT_EVENT_PRE_COMMIT. However that means if another code block
changed shard state to something else (e.g. FILE_TO_DELETE) before XACT_EVENT_PRE_COMMIT
we overwrite that. To prevent that problem, in case of failure we started to change
shard state, only if its current shard state is FILE_FINALIZED.
This UDF returns a shard placement from cache given shard id and placement id. At the
moment it iterates over all shard placements of given shard by ShardPlacementList and
searches given placement id in that list, which is not a good solution performance-wise.
However, currently, this function will be used only when there is a failed transaction.
If a need arises we can optimize this function in the future.
All router, real-time, task-tracker plannable queries should now have
full prepared statement support (and even use router when possible),
unless they don't go through the custom plan interface (which
basically just affects LANGUAGE SQL (not plpgsql) functions).
This is achieved by forcing postgres' planner to always choose a
custom plan, by assigning very low costs to plans with bound
parameters (i.e. ones were the postgres planner replanned the query
upon EXECUTE with all parameter values provided), instead of the
generic one.
This requires some trickery, because for custom plans to work the
costs for a non-custom plan have to be known, which means we can't
error out when planning the generic plan. Instead we have to return a
"faux" plan, that'd trigger an error message if executed. But due to
the custom plan logic that plan will likely (unless called by an SQL
function, or because we can't support that query for some reason) not
be executed; instead the custom plan will be chosen.
So far router planner had encapsulated different functionality in
MultiRouterPlanCreate. Modifications always go through router, selects
sometimes. Modifications always error out if the query is unsupported,
selects return NULL. Especially the error handling is a problem for
the upcoming extension of prepared statement support.
Split MultiRouterPlanCreate into CreateRouterPlan and
CreateModifyPlan, and change them to not throw errors.
Instead errors are now reported by setting the new
MultiPlan->plannigError.
Callers of router planner functionality now have to throw errors
themselves if desired, but also can skip doing so.
This is a pre-requisite for expanding prepared statement support.
While touching all those lines, improve a number of error messages by
getting them closer to the postgres error message guidelines.
The name CreatePhysicalPlan() hasn't been accurate for a while, and
the split of work between multi_planner() and CreatePhysicalPlan()
doesn't seem perfect. So rename to CreateDistributedPlan() and move a
bit more logic in there.
It can be useful, e.g. in the upcoming prepared statement support, to
be able to return an error from a function that is not raised
immediately, but can later be thrown. That allows e.g. to attempt to
plan a statment using different methods and to create good error
messages in each planner, but to only error out after all planners
have been run.
To enable that create support for deferred error messages that can be
created (supporting errorcode, message, detail, hint) in one function,
and then thrown in different place.
Andres Freund
Metin Doslu
Jason Petersen
Marco Slot
velioglu
Burak Yucesoy
Onder Kalaci
Murat Tuncer
Brian Cloutier
Marco Slot
Burak Velioglu
Burak Velioglu
Eren Basak
Brian Cloutier