On 2021/05/12 19:19, Fujii Masao wrote:
> 
> 
> On 2021/05/11 18:46, Masahiro Ikeda wrote:
>>
>>
>> On 2021/05/11 16:44, Fujii Masao wrote:
>>>
>>>
>>> On 2021/04/28 9:10, Masahiro Ikeda wrote:
>>>>
>>>>
>>>> On 2021/04/27 21:56, Fujii Masao wrote:
>>>>>
>>>>>
>>>>> On 2021/04/26 10:11, Masahiro Ikeda wrote:
>>>>>>
>>>>>> First patch has only the changes for pg_stat_wal view.
>>>>>> ("v6-0001-performance-improvements-of-reporting-wal-stats-without-introducing-a-new-variable.patch")
>>>>>>
>>>>>>
>>>>>>
>>>>>
>>>>> +        pgWalUsage.wal_records == prevWalUsage.wal_records &&
>>>>> +        walStats.wal_write == 0 && walStats.wal_sync == 0 &&
>>>>>> WalStats.m_wal_write should be checked here instead of 
>>>>>> walStats.wal_write?
>>>>
>>>> Thanks! Yes, I'll fix it.
>>>
>>> Thanks!
>>
>> Thanks for your comments!
>> I fixed them.
> 
> Thanks for updating the patch!
> 
>      if ((pgStatTabList == NULL || pgStatTabList->tsa_used == 0) &&
>          pgStatXactCommit == 0 && pgStatXactRollback == 0 &&
> +        pgWalUsage.wal_records == prevWalUsage.wal_records &&
> +        WalStats.m_wal_write == 0 && WalStats.m_wal_sync == 0 &&
> 
> I'm just wondering if the above WAL activity counters need to be checked.
> Maybe it's not necessary because "pgStatXactCommit == 0 && pgStatXactRollback
> == 0"
> is checked? IOW, is there really the case where WAL activity counters are 
> updated
> even when both pgStatXactCommit and pgStatXactRollback are zero?

Thanks for checking.

I haven't found the case yet. But, I added the condition because there is a
discussion that it's safer.

(It seems the mail thread chain is broken, Sorry...)
I copy the discussion at the time below.

https://www.postgresql.org/message-id/20210330.172843.267174731834281075.horikyota.ntt%40gmail.com
>>>> 3) Doing if (memcmp(&WalStats, &all_zeroes, sizeof(PgStat_MsgWal)) == 0)
>>>>    just to figure out if there's been any changes isn't all that
>>>>    cheap. This is regularly exercised in read-only workloads too. Seems
>>>>    adding a boolean WalStats.have_pending = true or such would be
>>>>    better.
>>>> 4) For plain backends pgstat_report_wal() is called by
>>>>    pgstat_report_stat() - but it is not checked as part of the "Don't
>>>>    expend a clock check if nothing to do" check at the top.  It's
>>>>    probably rare to have pending wal stats without also passing one of
>>>>    the other conditions, but ...
>>>
>>> I added the logic to check if the stats counters are updated or not in
>>> pgstat_report_stat() using not only generated wal record but also write/sync
>>> counters, and it can skip to call reporting function.
>>
>> I removed the checking code whether the wal stats counters were updated or 
>> not
>> in pgstat_report_stat() since I couldn't understand the valid case yet.
>> pgstat_report_stat() is called by backends when the transaction is finished.
>> This means that the condition seems always pass.
>
> Doesn't the same holds for all other counters?  If you are saying that
> "wal counters should be zero if all other stats counters are zero", we
> need an assertion to check that and a comment to explain that.
>
> I personally find it safer to add the WAL-stats condition to the
> fast-return check, rather than addin such assertion.


> +    if (pgWalUsage.wal_records != prevWalUsage.wal_records)
> +    {
> +        WalUsage    walusage;
> +
> +        /*
> +         * Calculate how much WAL usage counters were increased by 
> substracting
> +         * the previous counters from the current ones. Fill the results in
> +         * WAL stats message.
> +         */
> +        MemSet(&walusage, 0, sizeof(WalUsage));
> +        WalUsageAccumDiff(&walusage, &pgWalUsage, &prevWalUsage);
> 
> Isn't it better to move the code "prevWalUsage = pgWalUsage" into here?
> Because it's necessary only when pgWalUsage.wal_records !=
> prevWalUsage.wal_records.

Yes, I fixed it.


Regards,
-- 
Masahiro Ikeda
NTT DATA CORPORATION
diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c
index 2b106d8473..355778028e 100644
--- a/src/backend/executor/instrument.c
+++ b/src/backend/executor/instrument.c
@@ -17,6 +17,12 @@
 
 #include "executor/instrument.h"
 
+/*
+ * Buffer and generated WAL usage counters.
+ *
+ * The counters are accumulated values. There are infrastructures
+ * to add the values and calculate the difference within a specific period.
+ */
 BufferUsage pgBufferUsage;
 static BufferUsage save_pgBufferUsage;
 WalUsage	pgWalUsage;
diff --git a/src/backend/postmaster/checkpointer.c b/src/backend/postmaster/checkpointer.c
index e7e6a2a459..1761694a5b 100644
--- a/src/backend/postmaster/checkpointer.c
+++ b/src/backend/postmaster/checkpointer.c
@@ -505,7 +505,7 @@ CheckpointerMain(void)
 		pgstat_send_bgwriter();
 
 		/* Send WAL statistics to the stats collector. */
-		pgstat_report_wal();
+		pgstat_send_wal(true);
 
 		/*
 		 * If any checkpoint flags have been set, redo the loop to handle the
@@ -583,7 +583,7 @@ HandleCheckpointerInterrupts(void)
 		BgWriterStats.m_requested_checkpoints++;
 		ShutdownXLOG(0, 0);
 		pgstat_send_bgwriter();
-		pgstat_report_wal();
+		pgstat_send_wal(true);
 
 		/* Normal exit from the checkpointer is here */
 		proc_exit(0);			/* done */
diff --git a/src/backend/postmaster/pgstat.c b/src/backend/postmaster/pgstat.c
index e94f5f55c7..b65bcd5c9b 100644
--- a/src/backend/postmaster/pgstat.c
+++ b/src/backend/postmaster/pgstat.c
@@ -133,8 +133,8 @@ PgStat_MsgWal WalStats;
 
 /*
  * WAL usage counters saved from pgWALUsage at the previous call to
- * pgstat_report_wal(). This is used to calculate how much WAL usage
- * happens between pgstat_report_wal() calls, by substracting
+ * pgstat_send_wal(). This is used to calculate how much WAL usage
+ * happens between pgstat_send_wal() calls, by substracting
  * the previous counters from the current ones.
  */
 static WalUsage prevWalUsage;
@@ -852,9 +852,19 @@ pgstat_report_stat(bool disconnect)
 	TabStatusArray *tsa;
 	int			i;
 
-	/* Don't expend a clock check if nothing to do */
+	/*
+	 * Don't expend a clock check if nothing to do.
+	 *
+	 * Note: regarding to WAL statistics counters, it's not enough to check
+	 * only whether the wal record is generated or not. If a read transaction
+	 * is executed, wal records aren't nomally generated (although HOT makes
+	 * wal records). But, just writes and syncs the wal data may happen when
+	 * to flush buffers.
+	 */
 	if ((pgStatTabList == NULL || pgStatTabList->tsa_used == 0) &&
 		pgStatXactCommit == 0 && pgStatXactRollback == 0 &&
+		pgWalUsage.wal_records == prevWalUsage.wal_records &&
+		WalStats.m_wal_write == 0 && WalStats.m_wal_sync == 0 &&
 		!have_function_stats && !disconnect)
 		return;
 
@@ -948,7 +958,7 @@ pgstat_report_stat(bool disconnect)
 	pgstat_send_funcstats();
 
 	/* Send WAL statistics */
-	pgstat_report_wal();
+	pgstat_send_wal(true);
 
 	/* Finally send SLRU statistics */
 	pgstat_send_slru();
@@ -2918,7 +2928,7 @@ void
 pgstat_initialize(void)
 {
 	/*
-	 * Initialize prevWalUsage with pgWalUsage so that pgstat_report_wal() can
+	 * Initialize prevWalUsage with pgWalUsage so that pgstat_send_wal() can
 	 * calculate how much pgWalUsage counters are increased by substracting
 	 * prevWalUsage from pgWalUsage.
 	 */
@@ -3030,44 +3040,6 @@ pgstat_send_bgwriter(void)
 	MemSet(&BgWriterStats, 0, sizeof(BgWriterStats));
 }
 
-/* ----------
- * pgstat_report_wal() -
- *
- * Calculate how much WAL usage counters are increased and send
- * WAL statistics to the collector.
- *
- * Must be called by processes that generate WAL.
- * ----------
- */
-void
-pgstat_report_wal(void)
-{
-	WalUsage	walusage;
-
-	/*
-	 * Calculate how much WAL usage counters are increased by substracting the
-	 * previous counters from the current ones. Fill the results in WAL stats
-	 * message.
-	 */
-	MemSet(&walusage, 0, sizeof(WalUsage));
-	WalUsageAccumDiff(&walusage, &pgWalUsage, &prevWalUsage);
-
-	WalStats.m_wal_records = walusage.wal_records;
-	WalStats.m_wal_fpi = walusage.wal_fpi;
-	WalStats.m_wal_bytes = walusage.wal_bytes;
-
-	/*
-	 * Send WAL stats message to the collector.
-	 */
-	if (!pgstat_send_wal(true))
-		return;
-
-	/*
-	 * Save the current counters for the subsequent calculation of WAL usage.
-	 */
-	prevWalUsage = pgWalUsage;
-}
-
 /* ----------
  * pgstat_send_wal() -
  *
@@ -3075,24 +3047,38 @@ pgstat_report_wal(void)
  *
  * If 'force' is not set, WAL stats message is only sent if enough time has
  * passed since last one was sent to reach PGSTAT_STAT_INTERVAL.
- *
- * Return true if the message is sent, and false otherwise.
  * ----------
  */
-bool
+void
 pgstat_send_wal(bool force)
 {
-	/* We assume this initializes to zeroes */
-	static const PgStat_MsgWal all_zeroes;
 	static TimestampTz sendTime = 0;
 
 	/*
-	 * This function can be called even if nothing at all has happened. In
-	 * this case, avoid sending a completely empty message to the stats
-	 * collector.
+	 * First, to check the WAL stats counters were updated.
+	 *
+	 * Even if the 'force' is true, we don't need to send the stats if the
+	 * counters were not updated.
+	 *
+	 * We can know whether the counters were updated or not to check only
+	 * three counters. So, for performance, we don't allocate another memory
+	 * spaces and check the all stats like pgstat_send_slru().
+	 *
+	 * The current 'wal_records' is the same as the previous one means that
+	 * wal records weren't generated. So, the counters of 'wal_fpi',
+	 * 'wal_bytes', 'm_wal_buffers_full' are not updated neither.
+	 *
+	 * It's not enough to check the number of generated wal records, for
+	 * example the walwriter may write/sync the WAL although it doesn't
+	 * generate wal records. 'm_wal_write' and 'm_wal_sync' are zero means the
+	 * counters of time spent are zero too.
 	 */
-	if (memcmp(&WalStats, &all_zeroes, sizeof(PgStat_MsgWal)) == 0)
-		return false;
+	if (pgWalUsage.wal_records == prevWalUsage.wal_records &&
+		WalStats.m_wal_write == 0 && WalStats.m_wal_sync == 0)
+	{
+		Assert(WalStats.m_wal_buffers_full == 0);
+		return;
+	}
 
 	if (!force)
 	{
@@ -3100,13 +3086,40 @@ pgstat_send_wal(bool force)
 
 		/*
 		 * Don't send a message unless it's been at least PGSTAT_STAT_INTERVAL
-		 * msec since we last sent one.
+		 * msec since we last sent one to avoid overloading the stats
+		 * collector.
 		 */
 		if (!TimestampDifferenceExceeds(sendTime, now, PGSTAT_STAT_INTERVAL))
-			return false;
+			return;
 		sendTime = now;
 	}
 
+	/*
+	 * Set the counters related to generated WAL data if the counters were
+	 * updated.
+	 */
+	if (pgWalUsage.wal_records != prevWalUsage.wal_records)
+	{
+		WalUsage	walusage;
+
+		/*
+		 * Calculate how much WAL usage counters were increased by substracting
+		 * the previous counters from the current ones. Fill the results in
+		 * WAL stats message.
+		 */
+		MemSet(&walusage, 0, sizeof(WalUsage));
+		WalUsageAccumDiff(&walusage, &pgWalUsage, &prevWalUsage);
+
+		WalStats.m_wal_records = walusage.wal_records;
+		WalStats.m_wal_fpi = walusage.wal_fpi;
+		WalStats.m_wal_bytes = walusage.wal_bytes;
+
+		/*
+		 * Save the current counters for the subsequent calculation of WAL usage.
+		 */
+		prevWalUsage = pgWalUsage;
+	}
+
 	/*
 	 * Prepare and send the message
 	 */
@@ -3117,8 +3130,6 @@ pgstat_send_wal(bool force)
 	 * Clear out the statistics buffer, so it can be re-used.
 	 */
 	MemSet(&WalStats, 0, sizeof(WalStats));
-
-	return true;
 }
 
 /* ----------
diff --git a/src/include/executor/instrument.h b/src/include/executor/instrument.h
index fc87eed4fb..31dbfc832a 100644
--- a/src/include/executor/instrument.h
+++ b/src/include/executor/instrument.h
@@ -16,6 +16,12 @@
 #include "portability/instr_time.h"
 
 
+/*
+ * The accumulated counters for buffer usage.
+ *
+ * The reason the counters are accumulated values is to avoid unexpected
+ * reset because many callers may use them.
+ */
 typedef struct BufferUsage
 {
 	int64		shared_blks_hit;	/* # of shared buffer hits */
@@ -32,6 +38,15 @@ typedef struct BufferUsage
 	instr_time	blk_write_time; /* time spent writing */
 } BufferUsage;
 
+/*
+ * The accumulated counters for generated WAL usage.
+ *
+ * The reason the counters are accumulated values is the same as BufferUsage's one.
+ * And the reason to store only generated WAL usage and doesn't store WAL I/O
+ * activity, is that this is assumed for knowing the WAL usage in per query or
+ * transaction. So, common resources for the cluster like WAL I/O activity is
+ * not stored.
+ */
 typedef struct WalUsage
 {
 	int64		wal_records;	/* # of WAL records produced */
diff --git a/src/include/pgstat.h b/src/include/pgstat.h
index 72ff4a06d6..7727e83455 100644
--- a/src/include/pgstat.h
+++ b/src/include/pgstat.h
@@ -1091,8 +1091,7 @@ extern void pgstat_twophase_postabort(TransactionId xid, uint16 info,
 
 extern void pgstat_send_archiver(const char *xlog, bool failed);
 extern void pgstat_send_bgwriter(void);
-extern void pgstat_report_wal(void);
-extern bool pgstat_send_wal(bool force);
+extern void pgstat_send_wal(bool force);
 
 /* ----------
  * Support functions for the SQL-callable functions to

Reply via email to