{WIP} Use SQLite as the file format for CDC streaming.

Author: dimitri

src/bin/pgcopydb/catalog.c

@@ -183,7 +183,17 @@ static char *sourceDBcreateDDLs[] = {
 	"create table sentinel("
 	"  id integer primary key check (id = 1), "
 	"  startpos pg_lsn, endpos pg_lsn, apply bool, "
-	" write_lsn pg_lsn, flush_lsn pg_lsn, replay_lsn pg_lsn)"
+	"  write_lsn pg_lsn, flush_lsn pg_lsn, "
+	"  transform_lsn pg_lsn, "
+	"  replay_lsn pg_lsn)",
+
+	"create table cdc_files("
+	"  id integer primary key, filename text unique, timeline integer, "
+	"  startpos pg_lsn, endpos pg_lsn, "
+	"  start_time_epoch integer, done_time_epoch integer)",
+
+	"create table timeline_history("
+	"  tli integer primary key, startpos pg_lsn, endpos pg_lsn)"
 };
 
 
@@ -407,6 +417,28 @@ static char *targetDBcreateDDLs[] = {
 };
 
 
+static char *replayDBcreateDDLs[] = {
+	"create table output("
+	"  id integer primary key, "
+	"  action text, xid integer, lsn pg_lsn, timestamp text, "
+	"  message text)",
+
+	"create unique index o_a_lsn on output(action, lsn)",
+	"create unique index o_a_xid on output(action, xid)",
+
+	"create table stmt(hash text primary key, sql text)",
+	"create unique index stmt_hash on stmt(hash)",
+
+	"create table replay("
+	"  id integer primary key, "
+	"  action text, xid integer, lsn pg_lsn, timestamp text, "
+	"  stmt_hash text references stmt(hash), stmt_args jsonb)",
+
+	"create index r_xid on replay(xid)",
+	"create index r_lsn on replay(lsn)",
+};
+
+
 static char *sourceDBdropDDLs[] = {
 	"drop table if exists setup",
 	"drop table if exists section",
@@ -473,6 +505,12 @@ static char *targetDBdropDDLs[] = {
 };
 
 
+static char *replayDBdropDDLs[] = {
+	"drop table if exists output",
+	"drop table if exists replay"
+};
+
+
 /*
  * catalog_init_from_specs initializes our internal catalog database file from
  * a specification.
@@ -945,6 +983,13 @@ catalog_create_schema(DatabaseCatalog *catalog)
 			break;
 		}
 
+		case DATABASE_CATALOG_TYPE_REPLAY:
+		{
+			createDDLs = replayDBcreateDDLs;
+			count = sizeof(replayDBcreateDDLs) / sizeof(replayDBcreateDDLs[0]);
+			break;
+		}
+
 		default:
 		{
 			log_error("BUG: called catalog_init for unknown type %d",
@@ -1005,6 +1050,13 @@ catalog_drop_schema(DatabaseCatalog *catalog)
 			break;
 		}
 
+		case DATABASE_CATALOG_TYPE_REPLAY:
+		{
+			dropDDLs = replayDBdropDDLs;
+			count = sizeof(replayDBdropDDLs) / sizeof(replayDBdropDDLs[0]);
+			break;
+		}
+
 		default:
 		{
 			log_error("BUG: called catalog_drop_schema for unknown type %d",
@@ -7996,6 +8048,27 @@ catalog_bind_parameters(sqlite3 *db,
 
 		switch (p->type)
 		{
+			case BIND_PARAMETER_TYPE_NULL:
+			{
+				int rc = sqlite3_bind_null(ppStmt, n);
+
+				if (rc != SQLITE_OK)
+				{
+					log_error("[SQLite %d] Failed to bind \"%s\" to NULL: %s",
+							  rc,
+							  p->name,
+							  sqlite3_errstr(rc));
+					return false;
+				}
+
+				if (logSQL)
+				{
+					appendPQExpBuffer(debugParameters, "%s", "null");
+				}
+
+				break;
+			}
+
 			case BIND_PARAMETER_TYPE_INT:
 			{
 				int rc = sqlite3_bind_int(ppStmt, n, p->intVal);

src/bin/pgcopydb/catalog.h

@@ -628,6 +628,7 @@ bool catalog_count_summary_done_fetch(SQLiteQuery *query);
 typedef enum
 {
 	BIND_PARAMETER_TYPE_UNKNOWN = 0,
+	BIND_PARAMETER_TYPE_NULL,
 	BIND_PARAMETER_TYPE_INT,
 	BIND_PARAMETER_TYPE_INT64,
 	BIND_PARAMETER_TYPE_TEXT

src/bin/pgcopydb/cli_clone_follow.c

@@ -208,6 +208,7 @@ clone_and_follow(CopyDataSpec *copySpecs)
 						   copyDBoptions.endpos,
 						   STREAM_MODE_CATCHUP,
 						   &(copySpecs->catalogs.source),
+						   &(copySpecs->catalogs.replay),
 						   copyDBoptions.stdIn,
 						   copyDBoptions.stdOut,
 						   logSQL))
@@ -386,6 +387,7 @@ cli_follow(int argc, char **argv)
 						   copyDBoptions.endpos,
 						   STREAM_MODE_CATCHUP,
 						   &(copySpecs.catalogs.source),
+						   &(copySpecs.catalogs.replay),
 						   copyDBoptions.stdIn,
 						   copyDBoptions.stdOut,
 						   logSQL))

src/bin/pgcopydb/cli_snapshot.c

@@ -323,6 +323,7 @@ cli_create_snapshot(int argc, char **argv)
 							   createSNoptions.endpos,
 							   STREAM_MODE_CATCHUP,
 							   &(copySpecs.catalogs.source),
+							   &(copySpecs.catalogs.replay),
 							   createSNoptions.stdIn,
 							   createSNoptions.stdOut,
 							   logSQL))

src/bin/pgcopydb/cli_stream.c

@@ -582,6 +582,7 @@ cli_stream_setup(int argc, char **argv)
 						   streamDBoptions.endpos,
 						   STREAM_MODE_CATCHUP,
 						   &(copySpecs.catalogs.source),
+						   &(copySpecs.catalogs.replay),
 						   streamDBoptions.stdIn,
 						   streamDBoptions.stdOut,
 						   logSQL))
@@ -709,6 +710,7 @@ cli_stream_catchup(int argc, char **argv)
 						   streamDBoptions.endpos,
 						   STREAM_MODE_CATCHUP,
 						   &(copySpecs.catalogs.source),
+						   &(copySpecs.catalogs.replay),
 						   streamDBoptions.stdIn,
 						   streamDBoptions.stdOut,
 						   logSQL))
@@ -792,6 +794,7 @@ cli_stream_replay(int argc, char **argv)
 						   streamDBoptions.endpos,
 						   STREAM_MODE_REPLAY,
 						   &(copySpecs.catalogs.source),
+						   &(copySpecs.catalogs.replay),
 						   true,  /* stdin */
 						   true, /* stdout */
 						   logSQL))
@@ -800,17 +803,6 @@ cli_stream_replay(int argc, char **argv)
 		exit(EXIT_CODE_INTERNAL_ERROR);
 	}
 
-	/*
-	 * Remove the possibly still existing stream context files from
-	 * previous round of operations (--resume, etc). We want to make sure
-	 * that the catchup process reads the files created on this connection.
-	 */
-	if (!stream_cleanup_context(&specs))
-	{
-		/* errors have already been logged */
-		exit(EXIT_CODE_INTERNAL_ERROR);
-	}
-
 	/*
 	 * Before starting the receive, transform, and apply sub-processes, we need
 	 * to set the sentinel endpos to the command line --endpos option, when
@@ -917,6 +909,7 @@ cli_stream_transform(int argc, char **argv)
 						   streamDBoptions.endpos,
 						   STREAM_MODE_CATCHUP,
 						   &(copySpecs.catalogs.source),
+						   &(copySpecs.catalogs.replay),
 						   streamDBoptions.stdIn,
 						   streamDBoptions.stdOut,
 						   logSQL))
@@ -1079,6 +1072,7 @@ cli_stream_apply(int argc, char **argv)
 							   streamDBoptions.endpos,
 							   STREAM_MODE_CATCHUP,
 							   &(copySpecs.catalogs.source),
+							   &(copySpecs.catalogs.replay),
 							   true, /* streamDBoptions.stdIn */
 							   false, /* streamDBoptions.stdOut */
 							   logSQL))
@@ -1102,6 +1096,7 @@ cli_stream_apply(int argc, char **argv)
 
 		if (!stream_apply_init_context(&context,
 									   &(copySpecs.catalogs.source),
+									   &(copySpecs.catalogs.replay),
 									   &(copySpecs.cfPaths.cdc),
 									   &(streamDBoptions.connStrings),
 									   streamDBoptions.origin,
@@ -1190,6 +1185,7 @@ stream_start_in_mode(LogicalStreamMode mode)
 						   streamDBoptions.endpos,
 						   mode,
 						   &(copySpecs.catalogs.source),
+						   &(copySpecs.catalogs.replay),
 						   streamDBoptions.stdIn,
 						   streamDBoptions.stdOut,
 						   logSQL))
@@ -1214,18 +1210,6 @@ stream_start_in_mode(LogicalStreamMode mode)
 
 		case STREAM_MODE_PREFETCH:
 		{
-			/*
-			 * Remove the possibly still existing stream context files from
-			 * previous round of operations (--resume, etc). We want to make
-			 * sure that the catchup process reads the files created on this
-			 * connection.
-			 */
-			if (!stream_cleanup_context(&specs))
-			{
-				/* errors have already been logged */
-				exit(EXIT_CODE_INTERNAL_ERROR);
-			}
-
 			if (!followDB(&copySpecs, &specs))
 			{
 				/* errors have already been logged */

src/bin/pgcopydb/copydb.c

@@ -563,17 +563,21 @@ copydb_init_specs(CopyDataSpec *specs,
 	DatabaseCatalog *source = &(specs->catalogs.source);
 	DatabaseCatalog *filter = &(specs->catalogs.filter);
 	DatabaseCatalog *target = &(specs->catalogs.target);
+	DatabaseCatalog *replay = &(specs->catalogs.replay);
 
 	/* init the catalog type */
 	source->type = DATABASE_CATALOG_TYPE_SOURCE;
 	filter->type = DATABASE_CATALOG_TYPE_FILTER;
 	target->type = DATABASE_CATALOG_TYPE_TARGET;
+	replay->type = DATABASE_CATALOG_TYPE_REPLAY;
 
 	/* pick the dbfile from the specs */
 	strlcpy(source->dbfile, specs->cfPaths.sdbfile, sizeof(source->dbfile));
 	strlcpy(filter->dbfile, specs->cfPaths.fdbfile, sizeof(filter->dbfile));
 	strlcpy(target->dbfile, specs->cfPaths.tdbfile, sizeof(target->dbfile));
 
+	/* skip replay->dbfile which is rotated */
+
 	if (specs->section == DATA_SECTION_ALL ||
 		specs->section == DATA_SECTION_TABLE_DATA)
 	{

src/bin/pgcopydb/copydb.h

@@ -175,15 +175,16 @@ typedef struct ExtensionReqs
 
 /*
  * pgcopydb sentinel is a table that's created on the source catalog and allows
- * communicating elements from the outside, and in between the receive and
- * apply processes.
+ * communicating elements from the outside, and in between the receive,
+ * transform and apply processes.
  */
 typedef struct CopyDBSentinel
 {
 	bool apply;
 	uint64_t startpos;
 	uint64_t endpos;
 	uint64_t write_lsn;
+	uint64_t transform_lsn;
 	uint64_t flush_lsn;
 	uint64_t replay_lsn;
 } CopyDBSentinel;
@@ -491,6 +492,7 @@ bool sentinel_update_write_flush_lsn(DatabaseCatalog *catalog,
 									 uint64_t write_lsn,
 									 uint64_t flush_lsn);
 
+bool sentinel_update_transform_lsn(DatabaseCatalog *catalog, uint64_t transform_lsn);
 bool sentinel_update_replay_lsn(DatabaseCatalog *catalog, uint64_t replay_lsn);
 
 bool sentinel_get(DatabaseCatalog *catalog, CopyDBSentinel *sentinel);
@@ -501,6 +503,10 @@ bool sentinel_sync_recv(DatabaseCatalog *catalog,
 						uint64_t flush_lsn,
 						CopyDBSentinel *sentinel);
 
+bool sentinel_sync_transform(DatabaseCatalog *catalog,
+							 uint64_t transform_lsn,
+							 CopyDBSentinel *sentinel);
+
 bool sentinel_sync_apply(DatabaseCatalog *catalog,
 						 uint64_t replay_lsn,
 						 CopyDBSentinel *sentinel);

src/bin/pgcopydb/follow.c

@@ -251,18 +251,6 @@ follow_get_sentinel(StreamSpecs *specs, CopyDBSentinel *sentinel, bool verbose)
 bool
 follow_main_loop(CopyDataSpec *copySpecs, StreamSpecs *streamSpecs)
 {
-	/*
-	 * Remove the possibly still existing stream context files from
-	 * previous round of operations (--resume, etc). We want to make
-	 * sure that the catchup process reads the files created on this
-	 * connection.
-	 */
-	if (!stream_cleanup_context(streamSpecs))
-	{
-		/* errors have already been logged */
-		return false;
-	}
-
 	DatabaseCatalog *sourceDB = &(copySpecs->catalogs.source);
 
 	if (!catalog_open(sourceDB))
@@ -417,18 +405,6 @@ follow_prepare_mode_switch(StreamSpecs *streamSpecs,
 {
 	log_info("Catching-up from existing on-disk files");
 
-	if (streamSpecs->system.timeline == 0)
-	{
-		if (!stream_read_context(&(streamSpecs->paths),
-								 &(streamSpecs->system),
-								 &(streamSpecs->WalSegSz)))
-		{
-			log_error("Failed to read the streaming context information "
-					  "from the source database, see above for details");
-			return false;
-		}
-	}
-
 	/*
 	 * If the previous mode was catch-up, then before proceeding, we might need
 	 * to empty the transform queue where the STOP message was sent.
@@ -518,19 +494,6 @@ followDB(CopyDataSpec *copySpecs, StreamSpecs *streamSpecs)
 		return false;
 	}
 
-	/*
-	 * Before starting sub-processes, clean-up intermediate files from previous
-	 * round. Here that's the stream context with WAL segment size and timeline
-	 * history, which are fetched from the source server to compute WAL file
-	 * names. The current timeline can only change at a server restart or a
-	 * failover, both with trigger a reconnect.
-	 */
-	if (!stream_cleanup_context(streamSpecs))
-	{
-		/* errors have already been logged */
-		return false;
-	}
-
 	/*
 	 * Before starting sub-processes, make sure to close our SQLite catalogs.
 	 * We open the SQLite catalogs again before returning from this function