racket/collects/db/TODO
Ryan Culpepper f142a1c5f2 db: clean up communication, locking
Disconnect on break exn within lock; other break-safety fixes.
2012-02-08 22:32:51 -07:00

116 lines
4.1 KiB
Plaintext

----
Testing
- run ODBC tests on Mac
- run ODBC DB2, Oracle tests on win32; also test SQL Server ?
- test util/connect features
- test transaction functions
----
Types
- type annotations
- two modes: mandatory and opportunistic
- on result fields (eg sqlite, convert to date)
- on parameters ???
- per query or per connection? (or both?)
- either only well-known conversions, or must apply outside of lock
- postgresql record type: docs, send
- postgresql domain types, table record types, etc
- util/postgresql: add inet types
- util/geometry: add WKT functions
- add support for ODBC intervals (no point w/o driver to test with)
----
Misc
- internal docs
- use ffi/unsafe/alloc to simplify odbc handle allocation
- add ODBC-like functions for inspecting schemas (list-tables, etc)
- for wrapped/managed connections, detect if underlying connection gets
disconnected by server (eg, times out after 10 minutes of inactivity)
- at least, pool should make sure connection is alive when gotten from idle list
- add {keepalive : -> boolean} method to connection<%> (?)
- document/enumerate errors
- document exn:fail:sql, when used, when not used, links to SQLSTATE docs?
- disconnect on custudian shutdown (?)
- disconnect should always work, even on thread-damaged connections
- but might need version with timeout and/or rudely? flag, because
I can't think of a way to solve the lock problem that doesn't involve aux thread.
- finish transaction api: tests, custom sqlite options (?), custom mysql options (?)
- add connect option #:rollback-invalid-transactions? (?)
- identify common ODBC errors that can't possibly rollback transaction
- test with PostgreSQL ODBC driver using "rollback on all errors" mode
- more ODBC information (?)
SQLGetInfo:
- SQL_CONVERT_<datatype> - use to refine supported types (?)
- SQL_CURSOR_{COMMIT,ROLLBACK}_BEHAVIOR - check that commit/rollback doesn't delete pstmts!
- SQL_MAX_{CATALOG,COLUMN,IDENTIFIER,SCHEMA_NAME,TABLE_NAME}_NAME (min 128)
- add evt versions of functions
- for query functions (?)
- connection-pool-lease-evt
- when is it useful in practice?
- would make it easier to handle timeouts...
- on insert, return last inserted id
- postgresql: parse CommandComplete msg tag
- mysql: in ok-packet (what conditions, though?)
- sqlite3: sqlite3_last_insert_rowid(), use sqlite3_changes() to see if insert succeeded,
but still need to tell if stmt was even insert (parse sql?)
- odbc: ???
- add recursive locking?
- cons: - considered by experts to be bad design, sloppy
- pros: - would simplify cleanup for one-shot pstmts
- would enable simple impl of user-level 'call-with-lock' for grouping
multiple operations together
(but this could also be done by two locks: outer "ownership" lock
and inner "invariant-protecting" lock)
- audit code for break-safety, disable breaks as needed
- dialect info for ODBC
- can get some tx data from ODBC...
- on the other hand, not supposed to do tx-SQL in ODBC anyway, so low-priority
- make implementation notes section of docs
- explain cursor impl (& rationale)
- explain nested tx impl
- ...
- better query path: IMMEDIATE STATEMENTS vs PREPARED STATEMENT CACHE
- IMMEDIATE STATEMENTS
In pg terms, pipeline Parse and Bind (ie, 1 round-trip instead of 2).
Can do immediate query with args using typeids of parameters
- problem: results in worse error messages
- problem: error invalidates tx (in pg)
- problem: same problem exists even with no args: arity errors
- benefit: I like the code cleanup.
- PREPARED STATEMENT CACHE
Add sql=>pstmt cache in collection.
- schema can change without warning, invalidate pst type information
- mitigation: only cache w/in tx (but is isolation level sufficient?)
and clear cache on any tx rollback or potential ddl stmt
- problem: (cached) prepare and query not done atomically (two lock entries)
so a statement can be invalidated between cache retrieval and execution
- support logging
- audit: make sure no output within atomic section (in drracket, may thread switch)