dsa_allocate could not find 4 free pages

Hackers,

I'm getting the error in $subject and am only posting here because (a) the comments
in src/backend/utils/mmgr/dsa.c circa line 720 suggest that this is a bug, and (b) the
DSA code appears to be fairly new to the postgresql project, and perhaps not fully
debugged.

I am running against a build from 291a31c42c893574e9676e00121e6c6915a59de5
REL_10_STABLE, on centos linux.

I'm not asking for help configuring my system.  If this is a run of the mill out of memory
type error, I'll deal with it myself.  But if this indicates a bug, I'll be happy to try to distill
what I'm doing to a test case that I can share.  (I've only hit this once, and have little
idea whether it would be reproducible.)  The gist of the query was an array_agg over
the results of a parallel query over a partitioned table, executed from within a plpgsql
function, like:

SELECT ARRAY_AGG((a, latest, timelogged, b, cnt)::custom_composite_type) AS trans FROM
    (SELECT a, b, timelogged, MAX(timelogged) OVER (PARTITION BY a) AS latest, SUM(cnt) AS cnt FROM
        (SELECT a, b, timelogged, COUNT(*) AS cnt
            FROM mytable
            WHERE a= ANY('{65537,65538,65539,65540,65541,65542,65543,65544}'::OID[])
            AND timelogged >= '2017-12-04 16:12:50-08'::TIMESTAMP
            AND timelogged < '2017-12-04 17:12:50-08'::TIMESTAMP
            GROUP BY a, b, timelogged
        ) AS ss
        GROUP BY a, b, timelogged
) AS s

If this is not a bug, but rather something that users can hit through normal resource
exhaustion, then perhaps the comments in dsa.c can be changed to make that clear.

Thanks!

mark