Function DeleteExpiredRowsForIndexCore

pg_documentdb/src/ttl/ttl_index.c:602–845

Deletes the rows that have expired for the given table name and ttl entry information. * It deletes the number of items specified on the batchSize that have expired based on the index entry expiry value. * If ttlMetricsContext is not NULL, records metrics via the RecordTtlMetric hook. */

 * It deletes the number of items specified on the batchSize that have expired based on the index entry expiry value.
 * If ttlMetricsContext is not NULL, records metrics via the RecordTtlMetric hook. */
static uint64
DeleteExpiredRowsForIndexCore(char *tableName, TtlIndexEntry *indexEntry, int64
							  currentTime, int32 batchSize, instr_time startTime, int
							  budget, bool *isTaskTimeBudgetExceeded,
							  void *ttlMetricsContext)
{
	/* Per-batch timer for TTL deletion */
	instr_time batchStart;
	INSTR_TIME_SET_CURRENT(batchStart);

	int32 defaultBatchSize = (SkipRepeatDeleteForUnOrderedIndex &&
							  !indexEntry->indexIsOrdered) ?
							 MaxTTLBatchSizeUnorderedIndex : MaxTTLDeleteBatchSize;
	int32 ttlDeleteBatchSize = (batchSize != -1) ? batchSize :
							   defaultBatchSize;
	pgbson *indexKeyDocument = DatumGetPgBson(indexEntry->indexKeyDatum);
	pgbson *indexPfe = (indexEntry->indexPfeDatum != (Datum) 0) ?
					   DatumGetPgBson(indexEntry->indexPfeDatum) : NULL;

	/* TTL expireAfterSeconds is an int32 but we cast it as int64 to avoid overflow in milliseconds calculation */
	int64 indexExpiryMilliseconds = (int64) indexEntry->indexExpireAfterSeconds * 1000L;
	bson_iter_t pathSpecIter;
	PgbsonInitIterator(indexKeyDocument, &pathSpecIter);
	bson_iter_next(&pathSpecIter);
	const char *indexKey = bson_iter_key(&pathSpecIter);

	/*
	 *  Note that the TTL condition is repeated here, once for selecting a batch of records and once more while deleting.
	 *  The reason is that the SELECT subquery won't lock the records for update, so by the time we DELETE them, the records
	 *  could have been concurrently updated for them to come out of expiry window. Deleting those record won't be correct.
	 *  Hence we repeat the expiry condition once again in the DELETE statement.
	 *
	 *  As a side comment, we use the SELECT subquery to select a batch of records, because "DELETE" does not support LIMIT
	 *  (it's a Postgres limitation).
	 *
	 *  An index can have a partial filter expression (pfe) specified during creation which is also applicable for TTL index. We apply
	 *  the pfe predicate while selecting and deleting the expired records.
	 */

	StringInfo cmdStrDeleteRows = makeStringInfo();

	/* optimization for unsharded collection to avoid 2PC */
	appendStringInfo(cmdStrDeleteRows,
					 "DELETE FROM %s.%s"
					 " WHERE ctid IN (SELECT ctid FROM %s.%s"
					 " WHERE %s.bson_dollar_lt(document, $1::%s) ",
					 ApiDataSchemaName, tableName,
					 ApiDataSchemaName, tableName,
					 ApiCatalogSchemaName, FullBsonTypeName);

	int argCount = 1;

	/*
	 * In the SQL query above, we provide the corresponding TTL index as a hint for the TTL task query.
	 * Even though it's called a hint, by design it forces the use of the specified index.
	 *
	 *  As an additional note IndexOnly scans are not allowed for retrieving ctids.
	 */