citus/src/backend/distributed/worker/worker_shard_visibility.c

/*
 * worker_shard_visibility.c
 *
 * Implements the functions for hiding shards on the Citus MX
 * worker (data) nodes.
 *
 * Copyright (c) Citus Data, Inc.
 */

#include "postgres.h"

#include "catalog/index.h"
#include "catalog/namespace.h"
#include "catalog/pg_class.h"
#include "distributed/metadata_cache.h"
#include "distributed/coordinator_protocol.h"
#include "distributed/local_executor.h"
#include "distributed/worker_protocol.h"
#include "distributed/worker_shard_visibility.h"
#include "nodes/nodeFuncs.h"
#include "utils/lsyscache.h"
#include "utils/syscache.h"


/* Config variable managed via guc.c */
bool OverrideTableVisibility = true;
bool EnableManualChangesToShards = false;

static bool ReplaceTableVisibleFunctionWalker(Node *inputNode);

PG_FUNCTION_INFO_V1(citus_table_is_visible);
PG_FUNCTION_INFO_V1(relation_is_a_known_shard);


/*
 * relation_is_a_known_shard a wrapper around RelationIsAKnownShard(), so
 * see the details there. The function also treats the indexes on shards
 * as if they were shards.
 */
Datum
relation_is_a_known_shard(PG_FUNCTION_ARGS)
{
	CheckCitusVersion(ERROR);

	Oid relationId = PG_GETARG_OID(0);
	bool onlySearchPath = true;

	PG_RETURN_BOOL(RelationIsAKnownShard(relationId, onlySearchPath));
}


/*
 * citus_table_is_visible aims to behave exactly the same with
 * pg_table_is_visible with only one exception. The former one
 * returns false for the relations that are known to be shards.
 */
Datum
citus_table_is_visible(PG_FUNCTION_ARGS)
{
	CheckCitusVersion(ERROR);

	Oid relationId = PG_GETARG_OID(0);
	char relKind = '\0';
	bool onlySearchPath = true;

	/*
	 * We don't want to deal with not valid/existing relations
	 * as pg_table_is_visible does.
	 */
	if (!SearchSysCacheExists1(RELOID, ObjectIdGetDatum(relationId)))
	{
		PG_RETURN_NULL();
	}

	if (RelationIsAKnownShard(relationId, onlySearchPath))
	{
		/*
		 * If the input relation is an index we simply replace the
		 * relationId with the corresponding relation to hide indexes
		 * as well. See RelationIsAKnownShard() for the details and give
		 * more meaningful debug message here.
		 */
		relKind = get_rel_relkind(relationId);
		if (relKind == RELKIND_INDEX)
		{
			ereport(DEBUG2, (errmsg("skipping index \"%s\" since it belongs to a shard",
									get_rel_name(relationId))));
		}
		else
		{
			ereport(DEBUG2, (errmsg("skipping relation \"%s\" since it is a shard",
									get_rel_name(relationId))));
		}

		PG_RETURN_BOOL(false);
	}

	PG_RETURN_BOOL(RelationIsVisible(relationId));
}


/*
 * ErrorIfRelationIsAKnownShard errors out if the relation with relationId is
 * a shard relation.
 */
void
ErrorIfRelationIsAKnownShard(Oid relationId)
{
	/* search the relation in all schemas */
	bool onlySearchPath = false;
	if (!RelationIsAKnownShard(relationId, onlySearchPath))
	{
		return;
	}

	const char *relationName = get_rel_name(relationId);

	ereport(ERROR, (errmsg("relation \"%s\" is a shard relation ", relationName)));
}


/*
 * ErrorIfIllegallyChangingKnownShard errors out if the relation with relationId is
 * a known shard and manual changes on known shards are disabled. This is
 * valid for only non-citus (external) connections.
 */
void
ErrorIfIllegallyChangingKnownShard(Oid relationId)
{
	if (LocalExecutorLevel > 0 || IsCitusInitiatedRemoteBackend() ||
		EnableManualChangesToShards)
	{
		return;
	}

	/* search the relation in all schemas */
	bool onlySearchPath = true;
	if (RelationIsAKnownShard(relationId, onlySearchPath))
	{
		const char *relationName = get_rel_name(relationId);
		ereport(ERROR, (errmsg("cannot modify \"%s\" because it is a shard of "
							   "a distributed table",
							   relationName),
						errhint("Use the distributed table or set "
								"citus.enable_manual_changes_to_shards to on "
								"to modify shards directly")));
	}
}


/*
 * RelationIsAKnownShard gets a relationId, check whether it's a shard of
 * any distributed table. If onlySearchPath is true, then it searches
 * the current search path.
 *
 * We can only do that in MX since both the metadata and tables are only
 * present there.
 */
bool
RelationIsAKnownShard(Oid shardRelationId, bool onlySearchPath)
{
	bool missingOk = true;
	char relKind = '\0';

	if (!OidIsValid(shardRelationId))
	{
		/* we cannot continue without a valid Oid */
		return false;
	}

	if (IsCoordinator())
	{
		bool coordinatorIsKnown = false;
		PrimaryNodeForGroup(0, &coordinatorIsKnown);

		if (!coordinatorIsKnown)
		{
			/*
			 * We're not interested in shards in the coordinator
			 * or non-mx worker nodes, unless the coordinator is
			 * in pg_dist_node.
			 */
			return false;
		}
	}

	Relation relation = try_relation_open(shardRelationId, AccessShareLock);
	if (relation == NULL)
	{
		return false;
	}
	relation_close(relation, NoLock);

	/* we're not interested in the relations that are not in the search path */
	if (!RelationIsVisible(shardRelationId) && onlySearchPath)
	{
		return false;
	}

	/*
	 * If the input relation is an index we simply replace the
	 * relationId with the corresponding relation to hide indexes
	 * as well.
	 */
	relKind = get_rel_relkind(shardRelationId);
	if (relKind == RELKIND_INDEX)
	{
		shardRelationId = IndexGetRelation(shardRelationId, false);
	}

	/* get the shard's relation name */
	char *shardRelationName = get_rel_name(shardRelationId);

	uint64 shardId = ExtractShardIdFromTableName(shardRelationName, missingOk);
	if (shardId == INVALID_SHARD_ID)
	{
		/*
		 * The format of the table name does not align with
		 * our shard name definition.
		 */
		return false;
	}

	/* try to get the relation id */
	Oid relationId = LookupShardRelationFromCatalog(shardId, true);
	if (!OidIsValid(relationId))
	{
		/* there is no such relation */
		return false;
	}

	/* verify that their namespaces are the same */
	if (get_rel_namespace(shardRelationId) != get_rel_namespace(relationId))
	{
		return false;
	}

	/*
	 * Now get the relation name and append the shardId to it. We need
	 * to do that because otherwise a local table with a valid shardId
	 * appended to its name could be misleading.
	 */
	char *generatedRelationName = get_rel_name(relationId);
	AppendShardIdToName(&generatedRelationName, shardId);
	if (strncmp(shardRelationName, generatedRelationName, NAMEDATALEN) == 0)
	{
		/* we found the distributed table that the input shard belongs to */
		return true;
	}

	return false;
}


/*
 * ReplaceTableVisibleFunction is a wrapper around ReplaceTableVisibleFunctionWalker.
 * The replace functionality can be enabled/disable via a GUC. This function also
 * ensures that the extension is loaded and the version is compatible.
 */
void
ReplaceTableVisibleFunction(Node *inputNode)
{
	if (!OverrideTableVisibility ||
		!CitusHasBeenLoaded() || !CheckCitusVersion(DEBUG2))
	{
		return;
	}

	ReplaceTableVisibleFunctionWalker(inputNode);
}


/*
 * ReplaceTableVisibleFunction replaces all occurences of
 * pg_catalog.pg_table_visible() to
 * pg_catalog.citus_table_visible() in the given input node.
 *
 * Note that the only difference between the functions is that
 * the latter filters the tables that are known to be shards on
 * Citus MX worker (data) nodes.
 *
 * Note that although the function mutates the input node, we
 * prefer to use query_tree_walker/expression_tree_walker over
 * their mutator equivalents. This is safe because we ensure that
 * the replaced function has the exact same input/output values with
 * its precedent.
 */
static bool
ReplaceTableVisibleFunctionWalker(Node *inputNode)
{
	if (inputNode == NULL)
	{
		return false;
	}

	if (IsA(inputNode, FuncExpr))
	{
		FuncExpr *functionToProcess = (FuncExpr *) inputNode;
		Oid functionId = functionToProcess->funcid;

		if (functionId == PgTableVisibleFuncId())
		{
			/*
			 * We simply update the function id of the FuncExpr for
			 * two reasons: (i) We don't want to interfere with the
			 * memory contexts so don't want to deal with allocating
			 * a new functionExpr (ii) We already know that both
			 * functions have the exact same signature.
			 */
			functionToProcess->funcid = CitusTableVisibleFuncId();

			/* although not very likely, we could have nested calls to pg_table_is_visible */
			return expression_tree_walker(inputNode, ReplaceTableVisibleFunctionWalker,
										  NULL);
		}
	}
	else if (IsA(inputNode, Query))
	{
		return query_tree_walker((Query *) inputNode, ReplaceTableVisibleFunctionWalker,
								 NULL, 0);
	}

	return expression_tree_walker(inputNode, ReplaceTableVisibleFunctionWalker, NULL);
}