Postgres 11+ crashes with segmentation fault 11 when using SET STORAGE EXTERNAL on a `geographic` column

[progm]

Postgres 11+ crashes with segmentation fault 11 when using SET STORAGE EXTERNAL on a geographic column.

Being more specific, the segmentation fault happens when you switch to storage external a table containing a certain amount of data.

How to reproduce with docker (I've tested on macOS):

$ docker run --name postgis -p 5432:5432 -e POSTGRES_HOST_AUTH_METHOD=trust -e POSTGRES_PASSWORD=password postgis/postgis:11-3.1
$ psql -U postgres -h 127.0.0.1
postgres=# CREATE TABLE areas (id serial, geog geography(Multipolygon, 4326));
postgres=# insert into areas (id, geog) select i, ST_GeomFromText('MULTIPOLYGON (((11.174038669077522 44.00998460379789, 11.169693264583247 43.979730968523256, 11.08178331755044 43.970801288841, 11.011641246672006 44.00374236407625, 11.040369198606385 44.03420070463533, 11.049434195204165 44.09024166231109, 11.093226216052033 44.09445494894457, 11.107143581001756 44.05024384632795, 11.133566054440559 44.03986705890006, 11.125261503629277 44.022849011558975, 11.174038669077522 44.00998460379789)))') as geog from generate_series(1,1000) i;
postgres=# CREATE INDEX testing_idx ON areas USING GIST (geog);
postgres=# ALTER TABLE areas ALTER COLUMN geog SET STORAGE EXTERNAL;
postgres=# VACUUM FULL ANALYZE areas;

However, this seems to happen also if you switch to "storage external" first and you INSERT the data later.

I've tried exactly the same example with postgres 10 with postgis 3.1 and it works correctly with no segmentation fault.

Maybe it's a regression with postgres 11+ (you can replicate it with 12 and 13 as well).

Let me know if you need any further information.

[pramsey]

Milestone renamed

[robe]

In prep for 3.1.3 release

[pramsey]

Thanks for researching the PgSQL version. I can replicate your result, but the stack trace doesn't provide any obvious clues, except to show the problem is in the rebuild of the index that gets triggered on a REINDEX or a VACUUM FULL or even an UPDATE area SET geom = geom. The failure is in the index building code, but annoyingly not always in the same place!

[pramsey]

Also to note, I've tried a number of times to replicate without PostGIS. Trying with a tsvector column for example, but no luck.

[robe]

@pramsey did you try with the PostgreSQL geometry types? I would assume that is a closer proxy to PostGIS.

[Darafei Praliaskouski <me@…>]

In c22680a/git:

Add news item for GH658

Closes #4912

[Darafei Praliaskouski <me@…>]

In c22680a/git:

Add news item for GH658

Closes #4912