sql: 'name' attribute of DOid datum is lost during spilling to disk

[yuzefovich]

If we force spilling to disk, then we have some failures for pg_catalog test file. Namely, we lose name attribute of DOid datums for regproc type. For example, for this query:

query OTOOOOOOO colnames
SELECT oid, typname, typinput, typoutput, typreceive, typsend, typmodin, typmodout, typanalyze
FROM pg_catalog.pg_type
ORDER BY oid
----
oid         typname                                typinput        typoutput        typreceive        typsend           typmodin  typmodout  typanalyze
16          bool                                   boolin          boolout          boolrecv          boolsend          0         0          0
...

typinput, typoutput, typreceive, typsend columns will be printed incorrectly and will be incorrectly sent across the wire.

The root cause is that we're only encoding / decoding the integer when writing to disk, so we end up with an unpopulated DOid.name field.

As a possible solution we could include name into encoding / decoding as well, but that means we might make some data already written into the database no longer readable. My guess is that spilling to disk doesn't occur often enough, so we're probably ok with keeping this incompatibility unaddressed.

Jira issue: CRDB-14176

[mgartner]

This causes inconsistencies with Postgres for tables with REG* columns:

CockroachDB v21.2.7

defaultdb> create table t (r regtype);
CREATE TABLE

defaultdb> insert into t values ('int4'::regtype), ('bool'::regtype);
INSERT 2

defaultdb> select * from t;
  r
------
  23
  16
(2 rows)

Postgres 14.2

marcus=# create table t (r regtype);
CREATE TABLE

marcus=# insert into t values ('int4'::regtype), ('bool'::regtype);
INSERT 0 2

marcus=# select * from t;
    r
---------
 integer
 boolean
(2 rows)