interrupt tests fail due to change in error message

[strk]

I get these failures, when not disabling interrupt tests:

-----------------------------------------------------------------------------
 regress/core/interrupt_buffer .. failed (diff expected obtained: /tmp/pgis_reg/test_122_diff)
-----------------------------------------------------------------------------
--- /home/src/postgis/postgis/regress/core/interrupt_buffer_expected    2023-02-18 09:11:00.360153532 +0100
+++ /tmp/pgis_reg/test_122_out  2023-04-08 00:00:07.769328100 +0200
@@ -1,3 +1,3 @@
-ERROR:  canceling statement due to statement timeout
+ERROR:  canceling statement due to user request
 buffer interrupted on time
 1|5
-----------------------------------------------------------------------------

And many other similar subtle changes.

Also in regress/core/interrupt_relate and in regress/core/interrupt where the error message is slightly different (liblwgeom does not even try to loop for too long in that case now):

regress/core/interrupt .. failed (diff expected obtained: /tmp/pgis_reg/test_120_diff)

---

--- /home/src/postgis/postgis/regress/core/interrupt_expected 2023-02-18 09:11:00.360153532 +0100 +++ /tmp/pgis_reg/test_120_out 2023-04-08 00:00:06.593328067 +0200 @@ -1,4 +1,3 @@ -NOTICE: liblwgeom code interrupted -ERROR: canceling statement due to statement timeout +NOTICE: .././../liblwgeom/ptarray.c:449 - ptarray_segmentize2d: Too many segments required (3.605551e+100)

segmentize interrupted on time 1|LINESTRING(0 0,2 0,4 0)

---

I'm using the "buildbots" component becuase we don't seem to have a "testsuite" component (we probably should). I don't think there's any buildbot currently running these interrupt tests.

[strk]

The statement timeout error comes from PostgreSQL itself:

src/backend/tcop/postgres.c:3368: errmsg("canceling statement due to statement timeout")));

I think the error changed with commit [17708aa3c2cd5079b2b26f1befad43ec30036aa6/git] when a ! strstr was changed to a strstr conditional in this block:

/* Return NULL on GEOS error
 *
 * Prints error message only if it was not for interruption, in which
 * case we let PostgreSQL deal with the error.
 */
#define HANDLE_GEOS_ERROR(label) \
  { \
    if (strstr(lwgeom_geos_errmsg, "InterruptedException")) \
      ereport(ERROR, \
        (errcode(ERRCODE_QUERY_CANCELED), errmsg("canceling statement due to user request"))); \
    else \
      lwpgerror("%s: %s", (label), lwgeom_geos_errmsg); \
    PG_RETURN_NULL(); \
  }

The commit referenced #4070

Honestly I don't know if we can distinguish "user request" from "timeout", but I suspect the PostgreSQL error code would know it. For sure the testcase uses a timeout, which may not be the same thing as a "user request" (^C).

In any case, that commit changed the error message without updating the tests, which ended up under the carpet when CI failed to consistently succeed with them (due to real-time clock not being predictable with CI).

[strk]

I think the key here is "let PostgreSQL deal with the error", meaning IFF the geos error message is an interruption, we should do that, as a geos interruption is supposed to ONLY happen upon request from PostgreSQL, be it a timeout or a user request.

So the ! strstr condition was correct, in that it ONLY reported the GEOS error message when the GEOS exception was NOT an interrupt request.

What caused #4070 isn't clear to me, maybe pg_cancel_backend() triggers events in some unexpected order.

[Sandro Santilli <strk@…>]

In 1d263dc2/git:

Let PostgreSQL write the interruption error message

Closes #5365
Reverts 17708aa3c2

[Sandro Santilli <strk@…>]

In 4d851fb/git:

Let PostgreSQL write the interruption error message

References #5365 in stable-3.3 branch (3.3.4dev)

Tweak segmentize interrupt test