Parse typed literals in the where clause, starting with atoms

[jessestimpson]

This started as "let me filter on an atom" and turned into a general pass on non-string types in the where clause, plus a few parser bugs that showed up along the way.

Typed literals

A quoted literal is still always a string. For anything else you annotate it, Postgres style (CAST(x AS type) works too):

where status = 'active'::atom
where inserted_at >= '2024-03-01 12:00:00'::timestamp   -- NaiveDateTime
where seen_at >= '2024-03-01T12:00:00Z'::timestamptz    -- DateTime, UTC
where birthday = '2024-03-01'::date
where opens_at < '09:30:00'::time

naive_datetime and utc_datetime also work, for anyone thinking in Ecto types. A timestamp literal with an offset is rejected instead of quietly dropping the offset. A timestamptz without one is taken as UTC. Adding a new type is a clause or two in Efsql.Types.

Parsing was the easy part. Getting the right rows back needed two more fixes:

• The executor compared values with plain </==. On date and time structs that compares fields, not time, so 2025-01-02 sorted before 2024-01-03. Filtering, IN and ORDER BY now go through Types.compare/2. The same function fixes numbers against Decimal columns: before, price < 100 never matched a decimal and price > 100 matched all of them.
• efsql queries without a schema, so the adapter encoded datetime params with term_to_binary and never matched the index keys. That's why the README used to say to query indexed datetimes by their string form. The planner now encodes dates and times the same way the indexer writes them.

Range queries on an indexed decimal still can't work, because the adapter keys decimals by term encoding rather than value. That limit is in the README.

IS NULL

is null, is not null, plus the Postgres isnull/notnull shorthands. x = null now gives an error pointing at IS NULL instead of a FunctionClauseError.

Working around the parser

The sql library hangs or misparses in a few places, so efsql rewrites the token stream before handing it over. Efsql.Parser.parse/1 is now the only way in. Cases handled:

• ... and x is not null never returned.
• Any column named after a reserved word (day, at, user, value, ...) hung on where day = 'x'. It also misparsed with in/like/between, and turned order by day desc into ascending. This is what the CI timeouts were.

The sql dependency

I tried hex first, but 0.5.0 is older than the commit we were already on, so this tracks main (f6097ff). That commit fixes a hang on and (...). It also adds a Postgres client with C NIFs, whose startup opens a dets file in priv and sleeps 100ms. We only use the lexer and parser, so sql now goes in included_applications: it's loaded but never started. Building still needs cc/make for the NIFs, which is the main argument for eventually vendoring the lexer and parser.

CI

New test.yml runs mix test on PRs and pushes to main against a real fdbserver, pinned to the same versions as release.yml. It's green.

🤖 Generated with Claude Code

https://claude.ai/code/session_01PFBZUrerge6bbWVCHpS5Ko