gpbackup incorrectly backups HyperLogLog statistics (#41)

Sometimes gprestore cannot restore such backup.
This happens if HyperLogLog statistics were collected for a column whose
type cannot be implicitly converted to text, for example INTEGER.
In this case, gpbackup backups the binary representation of such statistics,
converting them to a column type, which is incorrect, since HyperLogLog
statistics are stored in a binary representation of the BYTEA type.
(https://github.com/arenadata/gpdb/blob/030faa2e79cbeb42ab4c843f8c8475c322506990/src/backend/commands/analyze.c#L1284-L1305)

This patch corrects the described issue by explicitly
specifying the BYTEA type for HyperLogLog statistics.

In the test, the leaf partition was excluded from checking
the number of statistics, because the current version of gpbackup
does not backup statistics for leaf partitions.
Once this is corrected, this additional condition can be removed.

(cherry picked from commit 85384fa)

backup/statistics.go

@@ -165,7 +165,10 @@ func generateAttributeSlotsQuery7(attStat AttributeStatistic) string {
 			AnyValues(attStat.Values2, attStat.Type),
 			AnyValues(attStat.Values3, attStat.Type),
 			AnyValues(attStat.Values4, attStat.Type),
-			AnyValues(attStat.Values5, attStat.Type))
+			// Hyperloglog data structure for STATISTIC_KIND_HLL and
+			// STATISTIC_KIND_FULLHLL is converted into a bytea and
+			// always stored in last slot
+			AnyValues(attStat.Values5, "bytea"))
 	}
 	return attributeQuery
 }
@@ -233,7 +236,10 @@ func generateAttributeSlotsQuery6(attStat AttributeStatistic) string {
 			AnyValues(attStat.Values2, attStat.Type),
 			AnyValues(attStat.Values3, attStat.Type),
 			AnyValues(attStat.Values4, attStat.Type),
-			AnyValues(attStat.Values5, attStat.Type))
+			// Hyperloglog data structure for STATISTIC_KIND_HLL and
+			// STATISTIC_KIND_FULLHLL is converted into a bytea and
+			// always stored in last slot
+			AnyValues(attStat.Values5, "bytea"))
 	}
 	return attributeQuery
 }
@@ -288,7 +294,9 @@ func generateAttributeSlotsQuery4(attStat AttributeStatistic) string {
 			AnyValues(attStat.Values1, attStat.Type),
 			AnyValues(attStat.Values2, attStat.Type),
 			AnyValues(attStat.Values3, attStat.Type),
-			AnyValues(attStat.Values4, attStat.Type))
+			// Hyperloglog data structure for STATISTIC_KIND_HLL is
+			// converted into a bytea and always stored in last slot
+			AnyValues(attStat.Values4, "bytea"))
 	}
 	return attributeQuery
 }

end_to_end/end_to_end_suite_test.go

@@ -1604,6 +1604,35 @@ var _ = Describe("backup and restore end to end tests", func() {
 			// gpbackup before version 1.18.0 does not dump pg_class statistics correctly
 			skipIfOldBackupVersionBefore("1.18.0")
 
+			testhelper.AssertQueryRuns(backupConn, `
+				CREATE TABLE et (
+					id character varying(13),
+					flg smallint,
+					dttm timestamp without time zone,
+					src character varying(80)
+				) WITH (appendonly='true', orientation='row', compresstype=zstd, compresslevel='3') DISTRIBUTED BY (id);
+
+				CREATE TABLE pt (
+					id character varying(13),
+					flg smallint,
+					dttm timestamp without time zone,
+					src character varying(80)
+				) WITH (appendonly='true', orientation='row', compresstype=zstd, compresslevel='3') DISTRIBUTED BY (id) PARTITION BY LIST(src) (
+					PARTITION src_mdm VALUES('val') WITH (tablename='pt_1_prt_src_mdm', appendonly='true', orientation='row', compresstype=zstd, compresslevel='3' )
+				);
+
+				INSERT INTO pt(id, flg, dttm, src) VALUES (1, 1, now(), 'val');
+				INSERT INTO et(id, flg, dttm, src) VALUES (2, 2, now(), 'val');
+
+				ANALYZE pt;
+				ANALYZE et;
+				ANALYZE ROOTPARTITION pt;
+
+				ALTER TABLE pt EXCHANGE PARTITION src_mdm WITH TABLE et;
+			`)
+
+			defer testhelper.AssertQueryRuns(backupConn,
+				`DROP TABLE et CASCADE; DROP TABLE pt CASCADE;`)
 			outputBkp := gpbackup(gpbackupPath, backupHelperPath,
 				"--with-stats",
 				"--backup-dir", backupDir, "--single-backup-dir")
@@ -1625,7 +1654,7 @@ var _ = Describe("backup and restore end to end tests", func() {
 			assertPGClassStatsRestored(backupConn, restoreConn, publicSchemaTupleCounts)
 			assertPGClassStatsRestored(backupConn, restoreConn, schema2TupleCounts)
 
-			statsQuery := fmt.Sprintf(`SELECT count(*) AS string FROM pg_statistic st left join pg_class cl on st.starelid = cl.oid left join pg_namespace nm on cl.relnamespace = nm.oid where %s;`, backup.SchemaFilterClause("nm"))
+			statsQuery := fmt.Sprintf(`SELECT count(*) AS string FROM pg_statistic st left join pg_class cl on st.starelid = cl.oid left join pg_namespace nm on cl.relnamespace = nm.oid where cl.relname != 'pt_1_prt_src_mdm' AND %s;`, backup.SchemaFilterClause("nm"))
 			backupStatisticCount := dbconn.MustSelectString(backupConn, statsQuery)
 			restoredStatisticsCount := dbconn.MustSelectString(restoreConn, statsQuery)