Detect column name collisions after PostgreSQL's 63-char identifier truncation

[dimitri]

Problem (fixes #353)

PostgreSQL limits identifier names to NAMEDATALEN-1 = 63 bytes. When two columns in the same table share the same first 63 characters they would both map to the same pg_attribute entry. Depending on timing this causes:

• CREATE TABLE to fail with column "x" of relation "t" specified more than once, or
• COPY to silently load data into the wrong column (if PostgreSQL truncates both names consistently and picks the first match).

pgloader never detected this case — it only forwarded the long names to PostgreSQL and let it fail opaquely. The original reporter (paracord55, 2016) could not reproduce the error; it has been confirmed again in 2026 with MSSQL → PostgreSQL migration.

The collision is per-table (PostgreSQL's uniqueness constraint for attname is (attrelid, attname)), not global — the same truncated name in two different tables is fine.

---

Fix

After the catalog has been fully transformed (cast rules applied, apply-identifier-case run, alter-table / alter-schema applied), but before any DDL or COPY is attempted:

1. Compute the effective PostgreSQL name for every column: strip outer double-quotes then take the first min(63, len) characters.
2. Group columns by effective name within each table.
3. Any group with more than one member is a collision.
4. Accumulate every instance across all tables, log them all at ERROR level (naming schema, table, the effective identifier, and the conflicting column names), then abort with a single fatal error that invites the user to rename the affected columns in the source database before migrating.

---

Changes

v3 (Common Lisp)

• src/utils/catalog.lisp — pg-effective-name, check-catalog-identifier-collisions
• src/package.lisp — export the new symbols
• src/load/migrate-database.lisp — call the check in process-catalog right after (cast catalog)

v4 (Clojure)

• clojure/src/pgloader/ddl/common.clj — pg-max-identifier-length (now public) + check-identifier-collisions
• clojure/src/pgloader/core.clj — call the check after the full catalog transformation pipeline, before truncate / DDL
• clojure/test/pgloader/ddl_test.clj — 7 new assertions (empty catalog, short names, exactly-63 names, no-collision with divergent prefixes, 2-column collision, 3-column collision, two-table collisions, no-cross-table-collision invariant)

Tests

• test/sqlite/create-collision.py + test/sqlite/collision.db — SQLite database with a products table whose two long column names share the same first 63 characters
• test/sqlite-collision.load — regression load file (in REGRESS)
• test/mysql-collision.load — manual fixture for MySQL (not in REGRESS; requires a running MySQL server)
• test/Makefile — sqlite-collision.load added to REGRESS with a special rule (LOAD DATABASE is incompatible with --regress)

MSSQL uses the same process-catalog / copy-database code path as MySQL and SQLite, so the fix covers MSSQL without a dedicated fixture (no SQL Server in local CI).

---

Example error output

LOG  Migrating from sqlite:///path/to/collision.db
LOG  Migrating into pgsql://localhost/pgloader
ERROR public.products: column name collision — "col_very_long..._aaax", "col_very_long..._aaay" all truncate to "col_very_long..._aaa"
FATAL Failed to process catalogs: 1 column name collision found in source catalog.

PostgreSQL limits identifier names to 63 bytes (NAMEDATALEN-1). The tables
listed above contain multiple columns whose names become identical after
truncation. This would cause CREATE TABLE to fail or COPY to load data into
the wrong column.

Please rename the affected columns in the source database before migrating.