SQL Classifier — midwess.com

PgPaw is intentionally read-only. It accepts one PostgreSQL SELECT statement over tables that exist in the local replica.

Accepted shape

select id, email
from users
where id = 7;

Joins are allowed:

select u.id, u.email, count(o.id) as order_count
from users u
left join orders o on o.user_id = u.id
where u.id = 7
group by u.id, u.email;

PgPaw rejects:

SQL parse failures;
multiple statements;
writes: insert, update, delete, merge;
DDL: create, alter, drop;
select into;
locking reads: for update, for share;
references to tables outside the replicated publication;
side-effecting functions, including nextval, setval, and advisory lock functions;
volatile reads, including now, random, random_normal, clock_timestamp, timeofday, statement_timestamp, gen_random_uuid, and uuid_generate_v4.

Rejected query example:

select nextval('orders_id_seq');

Response:

{
  "name": "RejectedError",
  "message": "function `nextval` has side effects and cannot be cached"
}

For accepted SQL, PgPaw records:

Field	Meaning
`fingerprint`	Opaque implementation fingerprint for the parsed statement.
`tables`	Lowercase table names, excluding CTE aliases.
`eq_filters`	Literal equality filters from the outer `WHERE` expression.
`sql`	Original SQL string used for execution.

The fingerprint appears in logs and snapshot URLs, but clients should treat it as opaque.

PgPaw can narrow cache invalidation when the query has a literal equality filter:

where id = 7

For a single-table query:

equality on the primary key anchors the version to that row;
equality on any replicated column can anchor if the table uses REPLICA IDENTITY FULL;
otherwise the query falls back to table-level invalidation.

Broad queries and joins use table-level invalidation.

Include primary key columns in live queries when possible.
Use deterministic SQL. Avoid time and random functions.
Keep private data protected with Postgres RLS and grants, not application-side filtering after PgPaw.
If a table is missing, check the publication first.