monitoring database activity database activity monitoring A database administrator frequently wonders, What is the system doing right now? This chapter discusses how to find that out. Several tools are available for monitoring database activity and analyzing performance. Most of this chapter is devoted to describing PostgreSQL's statistics collector, but one should not neglect regular Unix monitoring programs such as ps, top, iostat, and vmstat. Also, once one has identified a poorly-performing query, further investigation might be needed using PostgreSQL's command. discusses EXPLAIN and other methods for understanding the behavior of an individual query. ps to monitor activity On most platforms, PostgreSQL modifies its command title as reported by ps, so that individual server processes can readily be identified. A sample display is $ ps auxww | grep ^postgres postgres 960 0.0 1.1 6104 1480 pts/1 SN 13:17 0:00 postgres -i postgres 963 0.0 1.1 7084 1472 pts/1 SN 13:17 0:00 postgres: writer process postgres 965 0.0 1.1 6152 1512 pts/1 SN 13:17 0:00 postgres: stats collector process postgres 998 0.0 2.3 6532 2992 pts/1 SN 13:18 0:00 postgres: tgl runbug 127.0.0.1 idle postgres 1003 0.0 2.4 6532 3128 pts/1 SN 13:19 0:00 postgres: tgl regression [local] SELECT waiting postgres 1016 0.1 2.4 6532 3080 pts/1 SN 13:19 0:00 postgres: tgl regression [local] idle in transaction (The appropriate invocation of ps varies across different platforms, as do the details of what is shown. This example is from a recent Linux system.) The first process listed here is the master server process. The command arguments shown for it are the same ones given when it was launched. The next two processes are background worker processes automatically launched by the master process. (The stats collector process will not be present if you have set the system not to start the statistics collector.) Each of the remaining processes is a server process handling one client connection. Each such process sets its command line display in the form postgres: user database host activity The user, database, and connection source host items remain the same for the life of the client connection, but the activity indicator changes. The activity can be idle (i.e., waiting for a client command), idle in transaction (waiting for client inside a BEGIN block), or a command type name such as SELECT. Also, waiting is attached if the server process is presently waiting on a lock held by another server process. In the above example we can infer that process 1003 is waiting for process 1016 to complete its transaction and thereby release some lock or other. If you have turned off then the activity indicator is not updated; the process title is set only once when a new process is launched. On some platforms this saves a useful amount of per-command overhead, on others it's insignificant. Solaris requires special handling. You must use /usr/ucb/ps, rather than /bin/ps. You also must use two w flags, not just one. In addition, your original invocation of the postgres command must have a shorter ps status display than that provided by each server process. If you fail to do all three things, the ps output for each server process will be the original postgres command line. statistics PostgreSQL's statistics collector is a subsystem that supports collection and reporting of information about server activity. Presently, the collector can count accesses to tables and indexes in both disk-block and individual-row terms. PostgreSQL also supports determining the exact command currently being executed by other server processes. This is an independent facility that can be enabled or disabled whether or not block-level and row-level statistics are being collected. Since collection of statistics adds some overhead to query execution, the system can be configured to collect or not collect information. This is controlled by configuration parameters that are normally set in postgresql.conf. (See for details about setting configuration parameters.) The parameter must be set to true for the statistics collector to be launched at all. This is the default and recommended setting, but it can be turned off if you have no interest in statistics and want to squeeze out every last drop of overhead. (The savings is likely to be small, however.) Note that this option cannot be changed while the server is running. The parameters and control how much information is actually sent to the collector and thus determine how much run-time overhead occurs. These respectively determine whether a server process tracks disk-block-level access statistics and row-level access statistics and sends these to the collector. Additionally, per-database transaction commit and abort statistics are collected if either of these parameters are set. The parameter enables monitoring of the current command being executed by any server process. The statistics collector subprocess need not be running to enable this feature. Normally these parameters are set in postgresql.conf so that they apply to all server processes, but it is possible to turn them on or off in individual sessions using the command. (To prevent ordinary users from hiding their activity from the administrator, only superusers are allowed to change these parameters with SET.) Since the parameters stats_block_level, and stats_row_level default to false, very few statistics are collected in the default configuration. Enabling either of these configuration variables will significantly increase the amount of useful data produced by the statistics facilities, at the expense of additional run-time overhead. Several predefined views, listed in , are available to show the results of statistics collection. Alternatively, one can build custom views using the underlying statistics functions. When using the statistics to monitor current activity, it is important to realize that the information does not update instantaneously. Each individual server process transmits new block and row access counts to the collector just before going idle; so a query or transaction still in progress does not affect the displayed totals. Also, the collector itself emits a new report at most once per PGSTAT_STAT_INTERVAL milliseconds (500 unless altered while building the server). So the displayed information lags behind actual activity. However, current-query information collected by stats_command_string is always up-to-date. Another important point is that when a server process is asked to display any of these statistics, it first fetches the most recent report emitted by the collector process and then continues to use this snapshot for all statistical views and functions until the end of its current transaction. So the statistics will appear not to change as long as you continue the current transaction. Similarly, information about the current queries of all processes is collected when any such information is first requested within a transaction, and the same information will be displayed throughout the transaction. This is a feature, not a bug, because it allows you to perform several queries on the statistics and correlate the results without worrying that the numbers are changing underneath you. But if you want to see new results with each query, be sure to do the queries outside any transaction block. View Name Description pg_stat_activity One row per server process, showing database OID, database name, process ID, user OID, user name, current query, query's waiting status, time at which the current transaction and current query began execution, time at which the process was started, and client's address and port number. The columns that report data on the current query are available unless the parameter stats_command_string has been turned off. Furthermore, these columns are only visible if the user examining the view is a superuser or the same as the user owning the process being reported on. pg_stat_database One row per database, showing database OID, database name, number of active server processes connected to that database, number of transactions committed and rolled back in that database, total disk blocks read, and total buffer hits (i.e., block read requests avoided by finding the block already in buffer cache). pg_stat_all_tables For each table in the current database (including TOAST tables), the table OID, schema and table name, number of sequential scans initiated, number of live rows fetched by sequential scans, number of index scans initiated (over all indexes belonging to the table), number of live rows fetched by index scans, numbers of row insertions, updates, and deletions, the last time the table was vacuumed manually, the last time it was vacuumed by the autovacuum daemon, the last time it was analyzed manually, and the last time it was analyzed by the autovacuum daemon. pg_stat_sys_tables Same as pg_stat_all_tables, except that only system tables are shown. pg_stat_user_tables Same as pg_stat_all_tables, except that only user tables are shown. pg_stat_all_indexes For each index in the current database, the table and index OID, schema, table and index name, number of index scans initiated on that index, number of index entries returned by index scans, and number of live table rows fetched by simple index scans using that index. pg_stat_sys_indexes Same as pg_stat_all_indexes, except that only indexes on system tables are shown. pg_stat_user_indexes Same as pg_stat_all_indexes, except that only indexes on user tables are shown. pg_statio_all_tables For each table in the current database (including TOAST tables), the table OID, schema and table name, number of disk blocks read from that table, number of buffer hits, numbers of disk blocks read and buffer hits in all indexes of that table, numbers of disk blocks read and buffer hits from that table's auxiliary TOAST table (if any), and numbers of disk blocks read and buffer hits for the TOAST table's index. pg_statio_sys_tables Same as pg_statio_all_tables, except that only system tables are shown. pg_statio_user_tables Same as pg_statio_all_tables, except that only user tables are shown. pg_statio_all_indexes For each index in the current database, the table and index OID, schema, table and index name, numbers of disk blocks read and buffer hits in that index. pg_statio_sys_indexes Same as pg_statio_all_indexes, except that only indexes on system tables are shown. pg_statio_user_indexes Same as pg_statio_all_indexes, except that only indexes on user tables are shown. pg_statio_all_sequences For each sequence object in the current database, the sequence OID, schema and sequence name, numbers of disk blocks read and buffer hits in that sequence. pg_statio_sys_sequences Same as pg_statio_all_sequences, except that only system sequences are shown. (Presently, no system sequences are defined, so this view is always empty.) pg_statio_user_sequences Same as pg_statio_all_sequences, except that only user sequences are shown. The per-index statistics are particularly useful to determine which indexes are being used and how effective they are. Beginning in PostgreSQL 8.1, indexes can be used either directly or via bitmap scans. In a bitmap scan the output of several indexes can be combined via AND or OR rules; so it is difficult to associate individual heap row fetches with specific indexes when a bitmap scan is used. Therefore, a bitmap scan increments the pg_stat_all_indexes.idx_tup_read count(s) for the index(es) it uses, and it increments the pg_stat_all_tables.idx_tup_fetch count for the table, but it does not affect pg_stat_all_indexes.idx_tup_fetch. Before PostgreSQL 8.1, the idx_tup_read and idx_tup_fetch counts were essentially always equal. Now they can be different even without considering bitmap scans, because idx_tup_read counts index entries retrieved from the index while idx_tup_fetch counts live rows fetched from the table; the latter will be less if any dead or not-yet-committed rows are fetched using the index. The pg_statio_ views are primarily useful to determine the effectiveness of the buffer cache. When the number of actual disk reads is much smaller than the number of buffer hits, then the cache is satisfying most read requests without invoking a kernel call. However, these statistics do not give the entire story: due to the way in which PostgreSQL handles disk I/O, data that is not in the PostgreSQL buffer cache might still reside in the kernel's I/O cache, and might therefore still be fetched without requiring a physical read. Users interested in obtaining more detailed information on PostgreSQL I/O behavior are advised to use the PostgreSQL statistics collector in combination with operating system utilities that allow insight into the kernel's handling of I/O. Other ways of looking at the statistics can be set up by writing queries that use the same underlying statistics access functions as these standard views do. These functions are listed in . The per-database access functions take a database OID as argument to identify which database to report on. The per-table and per-index functions take a table or index OID. (Note that only tables and indexes in the current database can be seen with these functions.) The per-server-process access functions take a server process number, which ranges from one to the number of currently active server processes. Function Return Type Description pg_stat_get_db_numbackends(oid) integer Number of active server processes for database pg_stat_get_db_xact_commit(oid) bigint Transactions committed in database pg_stat_get_db_xact_rollback(oid) bigint Transactions rolled back in database pg_stat_get_db_blocks_fetched(oid) bigint Number of disk block fetch requests for database pg_stat_get_db_blocks_hit(oid) bigint Number of disk block fetch requests found in cache for database pg_stat_get_numscans(oid) bigint Number of sequential scans done when argument is a table, or number of index scans done when argument is an index pg_stat_get_tuples_returned(oid) bigint Number of rows read by sequential scans when argument is a table, or number of index entries returned when argument is an index pg_stat_get_tuples_fetched(oid) bigint Number of table rows fetched by bitmap scans when argument is a table, or table rows fetched by simple index scans using the index when argument is an index pg_stat_get_tuples_inserted(oid) bigint Number of rows inserted into table pg_stat_get_tuples_updated(oid) bigint Number of rows updated in table pg_stat_get_tuples_deleted(oid) bigint Number of rows deleted from table pg_stat_get_blocks_fetched(oid) bigint Number of disk block fetch requests for table or index pg_stat_get_blocks_hit(oid) bigint Number of disk block requests found in cache for table or index pg_stat_get_last_vacuum_time(oid) timestamptz Time of the last vacuum initiated by the user on this table pg_stat_get_last_autovacuum_time(oid) timestamptz Time of the last vacuum initiated by the autovacuum daemon on this table pg_stat_get_last_analyze_time(oid) timestamptz Time of the last analyze initiated by the user on this table pg_stat_get_last_autoanalyze_time(oid) timestamptz Time of the last analyze initiated by the autovacuum daemon on this table pg_stat_get_backend_idset() setof integer Set of currently active server process numbers (from 1 to the number of active server processes). See usage example in the text pg_backend_pid() integer Process ID of the server process attached to the current session pg_stat_get_backend_pid(integer) integer Process ID of the given server process pg_stat_get_backend_dbid(integer) oid Database ID of the given server process pg_stat_get_backend_userid(integer) oid User ID of the given server process pg_stat_get_backend_activity(integer) text Active command of the given server process, but only if the current user is a superuser or the same user as that of the session being queried (and stats_command_string is on) pg_stat_get_backend_waiting(integer) boolean True if the given server process is waiting for a lock, but only if the current user is a superuser or the same user as that of the session being queried (and stats_command_string is on) pg_stat_get_backend_activity_start(integer) timestamp with time zone The time at which the given server process' currently executing query was started, but only if the current user is a superuser or the same user as that of the session being queried (and stats_command_string is on) pg_stat_get_backend_start(integer) timestamp with time zone The time at which the given server process was started, or null if the current user is not a superuser nor the same user as that of the session being queried pg_stat_get_backend_client_addr(integer) inet The IP address of the client connected to the given server process. Null if the connection is over a Unix domain socket. Also null if the current user is not a superuser nor the same user as that of the session being queried pg_stat_get_backend_client_port(integer) integer The IP port number of the client connected to the given server process. -1 if the connection is over a Unix domain socket. Null if the current user is not a superuser nor the same user as that of the session being queried pg_stat_reset() boolean Reset all block-level and row-level statistics to zero blocks_fetched minus blocks_hit gives the number of kernel read() calls issued for the table, index, or database; but the actual number of physical reads is usually lower due to kernel-level buffering. The function pg_stat_get_backend_idset provides a convenient way to generate one row for each active server process. For example, to show the PIDs and current queries of all server processes: SELECT pg_stat_get_backend_pid(s.backendid) AS procpid, pg_stat_get_backend_activity(s.backendid) AS current_query FROM (SELECT pg_stat_get_backend_idset() AS backendid) AS s; lock monitoring Another useful tool for monitoring database activity is the pg_locks system table. It allows the database administrator to view information about the outstanding locks in the lock manager. For example, this capability can be used to: View all the locks currently outstanding, all the locks on relations in a particular database, all the locks on a particular relation, or all the locks held by a particular PostgreSQL session. Determine the relation in the current database with the most ungranted locks (which might be a source of contention among database clients). Determine the effect of lock contention on overall database performance, as well as the extent to which contention varies with overall database traffic. Details of the pg_locks view appear in . For more information on locking and managing concurrency with PostgreSQL, refer to . DTrace PostgreSQL provides facilities to support dynamic tracing of the database server. This allows an external utility to be called at specific points in the code and thereby trace execution. Currently, this facility is primarily intended for use by database developers, as it requires substantial familiarity with the code. A number of trace points, often called probes, are already inserted into the source code. By default these probes are disabled, and the user needs to explicitly tell the configure script to make the probes available in PostgreSQL. Currently, only the DTrace utility is supported, which is only available on Solaris Express and Solaris 10+. It is expected that DTrace will be available in the future on FreeBSD and Mac OS X. Supporting other dynamic tracing utilities is theoretically possible by changing the definitions for the PG_TRACE macros in src/include/pg_trace.h. By default, trace points are disabled, so you will need to explicitly tell the configure script to make the probes available in PostgreSQL. To include DTrace support specify --enable-dtrace to configure. See for further information. A few standard trace points are provided in the source code (of course, more can be added as needed for a particular problem). These are shown in . Name Parameters Overview transaction__start (int transactionId) The start of a new transaction. transaction__commit (int transactionId) The successful completion of a transaction. transaction__abort (int transactionId) The unsuccessful completion of a transaction. lwlock__acquire (int lockid, int mode) An LWLock has been acquired. lwlock__release (int lockid, int mode) An LWLock has been released. lwlock__startwait (int lockid, int mode) An LWLock was not immediately available and a backend has begun to wait for the lock to become available. lwlock__endwait (int lockid, int mode) A backend has been released from its wait for an LWLock. lwlock__condacquire (int lockid, int mode) An LWLock was successfully acquired when the caller specified no waiting. lwlock__condacquire__fail (int lockid, int mode) An LWLock was not successfully acquired when the caller specified no waiting. lock__startwait (int locktag_field2, int lockmode) A request for a heavyweight lock (lmgr lock) has begun to wait because the lock is not available. lock__endwait (int locktag_field2, int lockmode) A request for a heavyweight lock (lmgr lock) has finished waiting (i.e., has acquired the lock). The example below shows a DTrace script for analyzing transaction counts on the system, as an alternative to snapshotting pg_stat_database before and after a performance test. #!/usr/sbin/dtrace -qs postgresql$1:::transaction-start { @start["Start"] = count(); self->ts = timestamp; } postgresql$1:::transaction-abort { @abort["Abort"] = count(); } postgresql$1:::transaction-commit /self->ts/ { @commit["Commit"] = count(); @time["Total time (ns)"] = sum(timestamp - self->ts); self->ts=0; } Note how the double underline in trace point names needs to be replaced by a hyphen when using D script. When executed, the example D script gives output such as: # ./txn_count.d `pgrep -n postgres` ^C Start 71 Commit 70 Total time (ns) 2312105013 You should remember that trace programs need to be carefully written and debugged prior to their use, otherwise the trace information collected might be meaningless. In most cases where problems are found it is the instrumentation that is at fault, not the underlying system. When discussing information found using dynamic tracing, be sure to enclose the script used to allow that too to be checked and discussed. New trace points can be defined within the code wherever the developer desires, though this will require a recompilation. A trace point can be inserted by using one of the trace macros. These are chosen according to how many variables will be made available for inspection at that trace point. Tracing the occurrence of an event can be achieved with a single line, using just the trace point name, e.g. PG_TRACE (my__new__trace__point); More complex trace points can be provided with one or more variables for inspection by the dynamic tracing utility by using the PG_TRACEn macro that corresponds to the number of parameters after the trace point name: PG_TRACE3 (my__complex__event, varX, varY, varZ); The definition of the transaction__start trace point is shown below: static void StartTransaction(void) { ... /* * generate a new transaction id */ s->transactionId = GetNewTransactionId(false); XactLockTableInsert(s->transactionId); PG_TRACE1(transaction__start, s->transactionId); ... } Note how the transaction ID is made available to the dynamic tracing utility. The dynamic tracing utility might require you to further define these trace points. For example, DTrace requires you to add new probes to the file src/backend/utils/probes.d as shown here: provider postgresql { ... probe transaction__start(int); ... }; You should take care that the data types specified for the probe arguments match the datatypes of the variables used in the PG_TRACE macro. This is not checked at compile time. You can check that your newly added trace point is available by recompiling, then running the new binary, and as root, executing a DTrace command such as: dtrace -l -n transaction-start