On Wed, Feb 15, 2017 at 12:53 PM, Haribabu Kommi <kommi.harib...@gmail.com> wrote: > > > On Wed, Feb 8, 2017 at 9:36 PM, Amit Kapila <amit.kapil...@gmail.com> wrote: >> >> On Tue, Feb 7, 2017 at 11:47 AM, Haribabu Kommi >> <kommi.harib...@gmail.com> wrote: >> > Hi Hackers, >> > >> > I just want to discuss adding of a new statistics view that provides >> > the information of wal writing details as follows >> > >> >> +1. I think it will be useful to observe WAL activity. > > > Thanks for your opinion. > >> > postgres=# \d pg_stat_wal_writer >> > View "pg_catalog.pg_stat_wal_writer" >> > Column | Type | Collation | Nullable >> > | >> > Default >> > >> > -----------------------+--------------------------+-----------+----------+--------- >> > num_backend_writes | bigint | | >> > | >> > num_total_writes | bigint | | | >> > num_blocks | bigint | | | >> > total_write_time | bigint| | | >> > stats_reset | timestamp with time zone | | >> > | >> > >> > The columns of the view are >> > 1. Total number of xlog writes that are called from the backend. >> > 2. Total number of xlog writes that are called from both backend >> > and background workers. (This column can be changed to just >> > display on the background writes). >> > 3. The number of the blocks that are written. >> > 4. Total write_time of the IO operation it took, this variable data is >> > filled only when the track_io_timing GUC is enabled. >> >> So, here is *write_time* the total time system has spent in WAL >> writing before the last reset? > > > total write_time spent in WAL writing "after" the last reset in > milliseconds. > >> I think there should be a separate column for write and sync time. >> > > Added. > >> >> > Or it is possible to integrate the new columns into the existing >> > pg_stat_bgwriter view also. >> > >> >> I feel separate view is better. > > > Ok. > > Following the sample out of the view after regress run. > > postgres=# select * from pg_stat_walwrites; > -[ RECORD 1 ]--+------------------------------ > backend_writes | 19092 > writes | 663 > write_blocks | 56116 > write_time | 0 > sync_time | 3064 > stats_reset | 2017-02-15 13:37:09.454314+11 > > Currently, writer, walwriter and checkpointer processes > are considered as background processes that can do > the wal write mainly.
I'm not sure if this categorization is good. You told that this view is useful to tune walwriter parameters at the top of this thread. If so, ISTM that the information about walwriter's activity should be separated from others. What about other processes which *can* write WAL, for example walsender (e.g., BASE_BACKUP can cause WAL record), startup process (e.g., end-of- recovery checkpoint) and logical replication worker (Not sure if it always works with synchronous_commit=off, though). There might be other processes which can write WAL. Why didn't you separate "write_blocks", "write_time" and "sync_time" per the process category, like "backend_writes" and "writes"? This view doesn't seem to take into consideration the WAL writing and flushing during creating new WAL segment file. I think that it's better to make this view report also the number of WAL pages which are written when wal_buffer is full. This information is useful to tune the size of wal_buffers. This was proposed by Nagayasu before. https://www.postgresql.org/message-id/4ff824f3.5090...@uptime.jp Regards, -- Fujii Masao -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
