Greenplum segment级问题的排查 - gp_session_role=utility (含csvlog日志格式解读)
背景
有些时候，我们在使用Greenplum时，你会发现明明连接数是够的，但是报错的时候segment连接满无法正确连接，因为对外的接口都是MASTER，如果发现这里问题如何排查？
00000 LOG: Failed connection to seg0
00000 STATEMENT: INSERT INTO gps_track(
58M01 LOG: (58M01) Master unable to connect to seg0 with options : FATAL: Sorry, too many clients already
58M01 STATEMENT: INSERT INTO gps_track(
00000 LOG: Failed connection to seg1
00000 STATEMENT: INSERT INTO gps_track(
58M01 LOG: (58M01) Master unable to connect to seg1 with options : FATAL: Sorry, too many clients already
58M01 STATEMENT: INSERT INTO gps_track(
00000 LOG: The previous session was reset because its gang was disconnected (session id = 318465). The new session id = 318471
00000 STATEMENT: INSERT INTO gps_track(
00000 LOG: FTS rescanned, get new component databases info.
00000 STATEMENT: INSERT INTO gps_track(
00000 LOG: Failed connection to seg0
00000 STATEMENT: INSERT INTO gps_track(
58M01 LOG: (58M01) Master unable to connect to seg0 with options : FATAL: Sorry, too many clients already
58M01 STATEMENT: INSERT INTO gps_track(
00000 LOG: Failed connection to seg1
00000 STATEMENT: INSERT INTO gps_track(
58M01 LOG: (58M01) Master unable to connect to seg1 with options : FATAL: Sorry, too many clients already
58M01 STATEMENT: INSERT INTO gps_track(
00000 LOG: The previous session was reset because its gang was disconnected (session id = 318471). The new session id = 318482
00000 STATEMENT: INSERT INTO gps_track(
00000 LOG: no primary segworker group allocated
00000 STATEMENT: INSERT INTO gps_track(
XX000 ERROR: (XX000) Unexpected internal error (cdbgang.c:1636)
XX000 LOCATION: allocateWriterGang, cdbgang.c:1636
XX000 STATEMENT: INSERT INTO gps_track(
XX000 Traceback 0: postgres: port INSERT() [0x9099af]
XX000 Traceback 1: postgres: port INSERT(elog_internalerror+0x20) [0x90cdc0]
XX000 Traceback 2: postgres: port INSERT(allocateWriterGang+0x248) [0x9d6cb8]
XX000 Traceback 3: postgres: port INSERT(AssignGangs+0x28b) [0x6b689b]
XX000 Traceback 4: postgres: port INSERT(ExecutorStart+0xe5f) [0x69e28f]
XX000 Traceback 5: postgres: port INSERT() [0x8304b7]
XX000 Traceback 6: postgres: port INSERT() [0x830c54]
XX000 Traceback 7: postgres: port INSERT(PortalRun+0x322) [0x8311c2]
XX000 Traceback 8: postgres: port INSERT() [0x82abc1]
XX000 Traceback 9: postgres: port INSERT(PostgresMain+0xbd8) [0x82d418]
XX000 Traceback 10: postgres: port INSERT() [0x7d3637]
XX000 Traceback 11: postgres: port INSERT(PostmasterMain+0xd2e) [0x7d514e]
XX000 Traceback 12: postgres: port INSERT(main+0x193) [0x730363]
XX000 Traceback 13: /lib64/libc.so.6(__libc_start_main+0xfd) [0x2b6682187d5d]
XX000 Traceback 14: postgres: port INSERT() [0x4d6389]
备库到底是什么状态？
例子
方法1
通过utility模式连接（前提是预留了足够多的super user connection = superuser_reserved_connections），允许直接连到SEGMENT节点。
PGOPTIONS='-c gp_session_role=utility' psql -h xxx -p xxx -U xxx xxx
到segment节点一探究竟。
1、当前系统中的会话，评判标准：不能超过当前segment节点配置的max_connections 否则就会报连接不足。
select * from pg_stat_activity;
2、2PC会话，评判标准：不能超过当前segment节点配置的max_prepared_transactions 否则就会报连接不足。
postgres=# select * from pg_prepared_xacts ;
transaction | gid | prepared | owner | database
-------------+-----+----------+-------+----------
(0 rows)
3、segment级的锁等待状态
《Greenplum segment级锁问题排查方法 - 阿里云HybridDB for PostgreSQL最佳实践》
4、segment级日志排查。
cd $PGDATA/pg_log
查看日志
5、查看日志时，注意对应session id，可以找出它在master, primary segment上的所有相关日志。
日志格式如下：
src/backend/postmaster/syslogger.c
/*
* syslogger_write_errordata
* Write the GpErrorData to the log.
*/
static void
syslogger_write_errordata(PipeProtoHeader *chunkHeader, GpErrorData *errorData, bool csv)
{
syslogger_append_current_timestamp(true);
/* username */
syslogger_write_str_with_comma(errorData->username, true, csv, true);
/* databasename */
syslogger_write_str_with_comma(errorData->databasename, true, csv, true);
/* Process id, thread id */
syslogger_write_int32(false, "p", chunkHeader->pid, true, true);
syslogger_write_int32(false, "th", chunkHeader->thid, true, true);
/* Remote host */
syslogger_write_str_with_comma(errorData->remote_host, true, csv, true);
/* Remote port */
syslogger_write_str_with_comma(errorData->remote_port, true, csv, true);
/* session start timestamp */
syslogger_append_timestamp(errorData->fix_fields.session_start_time, true, true);
/* Transaction id */
syslogger_write_int32(false, "", errorData->fix_fields.top_trans_id, true, true);
/* GPDB specific options. */
syslogger_write_int32(true, "con", errorData->fix_fields.gp_session_id, true, true);
syslogger_write_int32(true, "cmd", errorData->fix_fields.gp_command_count, true, true);
syslogger_write_int32(false, errorData->fix_fields.gp_is_primary == 't'? "seg" : "mir", errorData->fix_fields.gp_segment_id,
true, true);
syslogger_write_int32(true, "slice", errorData->fix_fields.slice_id, true, true);
syslogger_write_int32(true, "dx", errorData->fix_fields.dist_trans_id, true, true);
syslogger_write_int32(true, "x", errorData->fix_fields.local_trans_id, true, true);
syslogger_write_int32(true, "sx", errorData->fix_fields.subtrans_id, true, true);
/* error severity */
syslogger_write_str_with_comma(errorData->error_severity, true, csv, true);
/* sql state code */
syslogger_write_str_with_comma(errorData->sql_state, true, csv, true);
/* errmsg */
syslogger_write_str_with_comma(errorData->error_message, true, csv, true);
/* errdetail */
syslogger_write_str_with_comma(errorData->error_detail, true, csv, true);
/* errhint */
syslogger_write_str_with_comma(errorData->error_hint, true, csv, true);
/* internal query */
syslogger_write_str_with_comma(errorData->internal_query, true, csv, true);
/* internal query pos */
syslogger_write_int32(true, "", errorData->fix_fields.internal_query_pos, true, true);
/* err ctxt */
syslogger_write_str_with_comma(errorData->error_context, true, csv, true);
/* user query */
syslogger_write_str_with_comma(errorData->debug_query_string, true, csv, true);
/* cursor pos */
syslogger_write_int32(false, "", errorData->fix_fields.error_cursor_pos, true, true);
/* func name */
syslogger_write_str_with_comma(errorData->error_func_name, true, csv, true);
/* file name */
syslogger_write_str_with_comma(errorData->error_filename, true, csv, true);
/* line number */
syslogger_write_int32(true, "", errorData->fix_fields.error_fileline, true, true);
/* stack trace */
if (errorData->stacktrace != NULL)
{
if (csv)
{
write_syslogger_file_binary("\"", 1, LOG_DESTINATION_STDERR);
}
syslogger_write_str(errorData->stacktrace, strlen(errorData->stacktrace), true, csv);
if (csv)
{
write_syslogger_file_binary("\"", 1, LOG_DESTINATION_STDERR);
}
}
/* EOL */
write_syslogger_file_binary(LOG_EOL, strlen(LOG_EOL), LOG_DESTINATION_STDERR);
/*
* Send alerts when needed. The alerts are sent only by the master.
* If the alert is failed for whatever reason, log a message and continue.
*/
if (errorData->fix_fields.send_alert == 't' &&
Gp_entry_postmaster && Gp_role == GP_ROLE_DISPATCH)
{
PG_TRY();
{
send_alert(errorData);
}
PG_CATCH();
{
elog(LOG,"Failed to send alert.");
}
PG_END_TRY();
}
}
方法2
另一种方式，我们也能够使用gp_dist_random下发指令给SEGMENT节点，（通过SQL接口连接到master节点调用SQL即可）
《Greenplum通过gp_dist_random(‘gp_id’) 在所有节点调用某个函数》
参考
《Greenplum segment级锁问题排查方法 - 阿里云HybridDB for PostgreSQL最佳实践》
《Greenplum segment节点直接读写配置与性能》
《Use pgbouncer connect to GreenPlum’s segment node》
《Greenplum通过gp_dist_random(‘gp_id’) 在所有节点调用某个函数》
《Greenplum & PostgreSQL UPSERT udf 实现 - 2 batch批量模式》
《Greenplum & PostgreSQL UPSERT udf 实现 - 1 单行模式》