Fix contrib/auto_explain to not cause problems in parallel workers.
authorTom Lane <tgl@sss.pgh.pa.us>
Mon, 3 Jun 2019 22:06:04 +0000 (18:06 -0400)
committerTom Lane <tgl@sss.pgh.pa.us>
Mon, 3 Jun 2019 22:06:04 +0000 (18:06 -0400)
A parallel worker process should not be making any decisions of its
own about whether to auto-explain.  If the parent session process
passed down flags asking for instrumentation data, do that, otherwise
not.  Trying to enable instrumentation anyway leads to bugs like the
"could not find key N in shm TOC" failure reported in bug #15821
from Christian Hofstaedtler.

We can implement this cheaply by piggybacking on the existing logic
for not doing anything when we've chosen not to sample a statement.

While at it, clean up some tin-eared coding related to the sampling
feature, including an off-by-one error that meant that asking for 1.0
sampling rate didn't actually result in sampling every statement.

Although the specific case reported here only manifested in >= v11,
I believe that related misbehaviors can be demonstrated in any version
that has parallel query; and the off-by-one error is certainly there
back to 9.6 where that feature was added.  So back-patch to 9.6.

Discussion: https://postgr.es/m/15821-5eb422e980594075@postgresql.org

contrib/auto_explain/auto_explain.c

index c810a41a7991757b43ea5de092e4bf4d31ddad36..a9536c2de053ebad9c657b58108f17064bf1171c 100644 (file)
@@ -14,6 +14,7 @@
 
 #include <limits.h>
 
+#include "access/parallel.h"
 #include "commands/explain.h"
 #include "executor/instrument.h"
 #include "jit/jit.h"
@@ -59,19 +60,20 @@ static const struct config_enum_entry loglevel_options[] = {
 /* Current nesting depth of ExecutorRun calls */
 static int     nesting_level = 0;
 
+/* Is the current top-level query to be sampled? */
+static bool current_query_sampled = false;
+
+#define auto_explain_enabled() \
+       (auto_explain_log_min_duration >= 0 && \
+        (nesting_level == 0 || auto_explain_log_nested_statements) && \
+        current_query_sampled)
+
 /* Saved hook values in case of unload */
 static ExecutorStart_hook_type prev_ExecutorStart = NULL;
 static ExecutorRun_hook_type prev_ExecutorRun = NULL;
 static ExecutorFinish_hook_type prev_ExecutorFinish = NULL;
 static ExecutorEnd_hook_type prev_ExecutorEnd = NULL;
 
-/* Is the current query sampled, per backend */
-static bool current_query_sampled = true;
-
-#define auto_explain_enabled() \
-       (auto_explain_log_min_duration >= 0 && \
-        (nesting_level == 0 || auto_explain_log_nested_statements))
-
 void           _PG_init(void);
 void           _PG_fini(void);
 
@@ -249,14 +251,25 @@ static void
 explain_ExecutorStart(QueryDesc *queryDesc, int eflags)
 {
        /*
-        * For rate sampling, randomly choose top-level statement. Either all
-        * nested statements will be explained or none will.
+        * At the beginning of each top-level statement, decide whether we'll
+        * sample this statement.  If nested-statement explaining is enabled,
+        * either all nested statements will be explained or none will.
+        *
+        * When in a parallel worker, we should do nothing, which we can implement
+        * cheaply by pretending we decided not to sample the current statement.
+        * If EXPLAIN is active in the parent session, data will be collected and
+        * reported back to the parent, and it's no business of ours to interfere.
         */
-       if (auto_explain_log_min_duration >= 0 && nesting_level == 0)
-               current_query_sampled = (random() < auto_explain_sample_rate *
-                                                                MAX_RANDOM_VALUE);
+       if (nesting_level == 0)
+       {
+               if (auto_explain_log_min_duration >= 0 && !IsParallelWorker())
+                       current_query_sampled = (random() < auto_explain_sample_rate *
+                                                                        ((double) MAX_RANDOM_VALUE + 1));
+               else
+                       current_query_sampled = false;
+       }
 
-       if (auto_explain_enabled() && current_query_sampled)
+       if (auto_explain_enabled())
        {
                /* Enable per-node instrumentation iff log_analyze is required. */
                if (auto_explain_log_analyze && (eflags & EXEC_FLAG_EXPLAIN_ONLY) == 0)
@@ -275,7 +288,7 @@ explain_ExecutorStart(QueryDesc *queryDesc, int eflags)
        else
                standard_ExecutorStart(queryDesc, eflags);
 
-       if (auto_explain_enabled() && current_query_sampled)
+       if (auto_explain_enabled())
        {
                /*
                 * Set up to track total elapsed time in ExecutorRun.  Make sure the
@@ -346,7 +359,7 @@ explain_ExecutorFinish(QueryDesc *queryDesc)
 static void
 explain_ExecutorEnd(QueryDesc *queryDesc)
 {
-       if (queryDesc->totaltime && auto_explain_enabled() && current_query_sampled)
+       if (queryDesc->totaltime && auto_explain_enabled())
        {
                double          msec;