Update symbols to pgcolumnar

[jdatcmd]

Multiple competing projects use columnar for shared libraries etc... to keep a distinct namespace we should use pgcolumnar

[ChronicallyJD]

Measured the collision rather than reasoning about it, since Citus columnar happens to
be installed on the bench host alongside pgColumnar.

Four real symbol collisions today

Comparing exported (dynamic, defined) symbols of the two shared libraries:

$ comm -12 <(nm -D --defined-only pgcolumnar.so | awk '{print $3}' | sort -u) \
           <(nm -D --defined-only citus_columnar.so | awk '{print $3}' | sort -u)
Pg_magic_func
_PG_init
columnar_handler
columnar_relation_storageid
pg_finfo_columnar_handler
pg_finfo_columnar_relation_storageid

Pg_magic_func and _PG_init are on every extension and are expected. The other four
are ours and are avoidable: columnar_handler and columnar_relation_storageid,
plus their pg_finfo_ pairs. columnar_handler is the table access method handler, so
it is the worst one on the list to share a name for.

Scale of the rename

pgColumnar exports 63 symbols. 27 begin columnar_, and only 5 begin
pgcolumnar. The 27:

columnar_add_projection            columnar_parallel_copy
columnar_cluster                   columnar_parallel_export_parquet
columnar_compact                   columnar_parquet_schema
columnar_compact_rewrite           columnar_read_parquet
columnar_debug_advance_reserved_offset  columnar_read_projection
columnar_debug_encoding_selftest   columnar_recluster
columnar_debug_set_metapage_version     columnar_reconstruct_via_projection
columnar_drop_projection           columnar_relation_storageid
columnar_export_arrow              columnar_truncate
columnar_export_parquet            columnar_vacuum
columnar_file_split_offsets        columnar_vacuum_sorted
columnar_handler                   columnar_vm_is_visible
columnar_import_arrow              columnar_vm_selftest
columnar_import_parquet

Those are the SQL-callable entry points, so each rename is a C symbol and its
CREATE FUNCTION in pgcolumnar--1.0-dev.sql. They are user-visible: columnar_vacuum
is what pgcolumnar.vacuum resolves to.

A user-visible collision the symbol list does not show

EXPLAIN on a Citus columnar table prints:

Custom Scan (ColumnarScan) on cpu_citus

which is byte-identical to ours. I hit this while writing the cross-engine benchmark
and had to disambiguate by which cluster the query ran on, because the plan line alone
does not tell you which extension produced it. Worth renaming the CustomScanMethods
name alongside the C symbols, or the plans stay ambiguous even after the link-level
problem is gone.

One caveat on urgency

PostgreSQL loads modules with RTLD_GLOBAL on most platforms, so co-loading both is
where a duplicate definition could bind the wrong way. On this host they are in
separate clusters (pgcolumnar preloaded on 5888, citus_columnar on 5899), so I
have not observed an actual misbind, and I do not want to claim one I did not see.
The collision is real; the failure is currently hypothetical.

Happy to take this if you want it, though it is a wide diff touching the SQL file and
every callsite, so it is probably a post-alpha change rather than something to squeeze
into the cut.

[ChronicallyJD]

Taking this. Will scope it first and post what the diff actually has to touch before writing it, since the user-visible blast radius depends on whether the SQL names are already namespaced.

[ChronicallyJD]

Scoped it before writing the diff, as promised. Two things changed my plan.

The export surface is not the same on every major

Built author/main against each major in the audit container and read nm -D:

major	exported	columnar_*	Columnar*	fb_*	pg_finfo_*
15	257	52	145	24	29
16	63	27	0	0	29
17	63	27	0	0	29
18	63	27	0	0	29

194 symbols export on 15 that do not export on 16 and up. The reason is PGXS, not us.
-fvisibility=hidden is in Makefile.shlib from 16 onward and absent on 15:

$ grep -ho '\-fvisibility=[a-z]*' /usr/local/pg16a/lib/postgresql/pgxs/src/Makefile.*
-fvisibility=hidden
$ grep -ho '\-fvisibility=[a-z]*' /usr/local/pg15a/lib/postgresql/pgxs/src/Makefile.*
(nothing)

So my original count of 27 was the PG18 figure, and PG15 is the exposed build.

The latent collision is worse than the four I reported

Citus on the bench is a PG18 build, so it exports only 27 symbols and still overlaps us
on the same 4. Comparing our PG15 exports against Citus's full symbol table, which
is what the overlap would be if both were built where nothing is hidden, gives 12 of ours:

ColumnarBeginRead                 columnar_chunk_group_row_limit
ColumnarEndRead                   columnar_compression
ColumnarReadNextRow               columnar_compression_level
ColumnarWriteRow                  columnar_stripe_row_limit
columnar_handler                  pg_finfo_columnar_handler
columnar_relation_storageid       pg_finfo_columnar_relation_storageid

The four in the right column are the ones I had not considered. They are GUC backing
variables, not functions. Our GUC strings are already pgcolumnar.*, but the C
variables kept the columnar_* names, and Citus uses the identical names for the
identical settings. A duplicate definition there does not crash. It silently binds one
library's setting to the other's storage, which is the worst failure mode on the list.

The blast radius is much smaller than I implied

I said this was a wide diff touching the SQL file and every callsite. That was wrong.
The SQL names are already namespaced:

CREATE FUNCTION pgcolumnar.vacuum(tablename regclass, stripe_count int DEFAULT 0)
	...
	AS 'MODULE_PATHNAME', 'columnar_vacuum';

columnar_vacuum appears only as the link name, the second argument. Renaming the C
symbol changes that string and nothing a user types. The one exception is
pgcolumnar.columnar_handler, whose SQL name carries the prefix too and is referenced
by CREATE ACCESS METHOD.

What I plan to do

1. 27 SQL-callable columnar_* to pgcolumnar_*. Fixes the observed collision on every
   major. The 27 pg_finfo_ twins follow automatically.
2. 25 GUC backing variables columnar_* to pgcolumnar_*. Fixes the silent-bind case.
3. 145 Columnar* to PgColumnar*, and 24 fb_* to pgc_fb_*. Internal only. fb_init,
   fb_start, fb_end and fb_grow are generic enough to collide with anything.
4. The CustomScanMethods name, so EXPLAIN stops printing a line identical to Citus's.

One thing worth doing whether or not we rename

Add -fvisibility=hidden to the PG15 build. It removes 194 exported symbols at a stroke
and makes 15 behave like 16 and up. The rename is the namespace fix. This is the
structural one, and the two are independent.

Question before I write it: do you want item 4 in the same PR? It is the only
user-visible change in the set, it edits expected/ output, and it is the one a user
could notice. I would rather land 1 to 3 as a no-op-to-users PR and do 4 separately.