Skip to content

LIMITATIONS.md

Latest commit

 

History

History
126 lines (123 loc) · 8.44 KB

LIMITATIONS.md

File metadata and controls

126 lines (123 loc) · 8.44 KB
Raw

MiniSQL 1.0 known limitations

  • Windows x64 PE and Linux x64 ELF are native targets. The graphical Workbench remains Windows-only.
  • The portable Java 11+ JDBC driver implements core connections, statements, prepared statements, transactions, forward-only results, and catalog metadata. Protocol v1 carries result cells as text without SQL type or NULL metadata, so ordinary result metadata is conservative, the text token NULL is ambiguous, and setBytes is unavailable until the SQL grammar gains a binary literal or the wire protocol gains typed parameter binding. Callable statements, generated keys, scrollable/updatable results, XA, and JDBC savepoint objects are not implemented in driver 1.0.
  • The Python 3.10+ DB-API connector exposes forward-only string/None rows, because protocol v1 has no SQL type or null bitmap. Text containing exactly NULL is ambiguous, binary parameter binding is unavailable, and one connection can own only one unread result stream. Stored procedures, multiple result sets, scrollable cursors, and two-phase commit are not implemented.
  • Linux offline tools, TLS, and bounded concurrent server/client operation pass the portable acceptance gate. The original EAGAIN/EWOULDBLOCK multi-client failure is retained as historical evidence in tests/performance/WINDOWS_LINUX_COMPARISON_2026-08-26.md; it was resolved by the pthread-backed MiniLang runtime and EINTR/readiness handling in the socket facade. Linux performance still needs the same workload-specific qualification as Windows before a production rollout.
  • Native TLS uses Schannel/CryptoAPI on Windows and OpenSSL 3 on Linux. Linux server credentials currently require an unencrypted PEM certificate/key pair. The current strict profile supports TLS 1.3 with TLS_AES_256_GCM_SHA384 and X25519 only; peers offering only different suites or groups are rejected.
  • Replication is asynchronous and operator-managed.
  • External sort and partitioned hash operators reduce working-set memory. The configured runtime.temporaryMemoryBytes value supplies a soft per-session, per-query budget and EXPLAIN ANALYZE reports estimated peak and spill counters. Eligible simple single-table queries stream directly from a storage cursor; blocking and unsupported shapes plus the compatibility client API still materialize the final QueryResult array.
  • The cost optimizer uses row/page/null/distinct/width estimates, eight-bucket equi-depth histograms for compact numeric/date values, integral or hashed most-common values, plus joint distinct counts and tuple MCVs for composite index keys through eight columns. It does not yet persist full wide-value ordering, expression distributions, or functional-dependency statistics. Inner-equijoin ordering uses bounded left-deep dynamic programming through eight sources and a deterministic greedy fallback above that limit; bushy trees and outer-join reordering are not implemented. Index-only scans cover B+ tree key columns and explicit INCLUDE payload columns. Partial indexes support immutable base expressions. Predicate implication is deliberately bounded to identical typed conjuncts and stronger typed single-column literal bounds in single-table plans; equivalent general expressions and floating-point bounds are not yet proven. Top-level inner equality graphs propagate non-NULL constants, while outer-join inference is deliberately excluded. The current ON CONFLICT(column) syntax cannot name a partial predicate, so it does not infer a partial unique index. Functional indexes currently accept exactly one deterministic expression key; mixed/composite expression keys, expression statistics, expression-aware index-only decoding, and functional join probes are not implemented. The query expression must have an identical typed binding. Expression identifiers that lose their meaning without SQL quoting are rejected until expression ASTs preserve quoting in canonical catalog SQL. One encoded B+ tree key is limited to 256 bytes and one leaf payload to 3,584 bytes so an individual entry always fits the minimum supported 4 KiB database page. Execution uses bounded batches and streaming fast paths where supported, including filtered scalar aggregates, page-range parallel scalar aggregates, fused single-table Top-N, and the final edge of reordered inner-equijoin COUNT(*). Large eligible hash joins and grouped aggregates use disk-backed partitions on up to four native workers. Other blocking operators and the final result array may still materialize large row sets; protocol-v1 sends that array in bounded continuation frames. The query memory budget is a spill policy, not a hard process-memory cap.
  • Configured servers hard-limit live connections, SQL payload bytes, response frame bytes, aggregate result bytes, result rows, shared spill reservations, and idle lifetime. Statements have cooperative cancellation and an absolute deadline; scans poll per physical heap page and DML batches per row. This is not unsafe asynchronous thread termination, so one native call or one very large scalar expression can delay observation until it returns. The managed process-memory setting is enforced on admission and cooperative poll boundaries using live MiniLang heap bytes, not total OS RSS, native TLS/socket allocation, or memory maps.
  • Table, column, schema-extension, statistics, and authorization metadata are not constrained to one page or a fixed 1 MiB snapshot. Capacity is still finite at the storage format's integer representation, process address space, and available disk; protocol frames and individual values retain defensive bounds and are not schema-object-count limits.
  • Recursive CTEs support an anchor and one recursive UNION/UNION ALL term; explicit mutual recursion is not supported and a 10,000-iteration guard diagnoses non-terminating statements.
  • Correlated subqueries require explicitly qualified outer references and are supported in non-grouped projection/filter/order expressions, not join predicates or aggregate/window queries.
  • Derived tables require aliases. Schema names are two-part only and there is no configurable search path or cross-database catalog qualifier.
  • Window functions operate on the complete partition. Explicit ROWS, RANGE, GROUPS, frame bounds and named windows are not implemented.
  • Trigger and stored-procedure bodies contain one supported DML statement and have no procedural control language. OLD/NEW are read-only, trigger recursion is bounded, and body RETURNING, ON CONFLICT, and INSERT ... SELECT forms are rejected rather than partially persisted.
  • MERGE accepts one source table and one action per matched/not-matched branch; conditional branches, multiple matched clauses and a query source are not yet implemented.
  • DCL is autocommit-only.
  • Connections, framing, SQL parsing and read-only query plans are threaded. Multiple reads can execute against one database simultaneously; mutations, maintenance, sequence consumption and session-state changes are serialized by a writer-prioritized per-database gate. There is still only one active physical writer per database. Long-running readers can delay a waiting writer until the current reader set completes; new readers do not bypass that writer. This concurrency contract is validated on both native targets.
  • The bundled automatic HA controller is a single-host reference with a stable loopback endpoint and native lease fencing. Its file witness is not distributed consensus and must have exactly one writer plus atomic replacement semantics. Multi-host election, quorum commit, synchronous replication, and storage-level fencing remain deployment responsibilities. Cross-database transactions are not implemented.
  • The M48 WAL stream covers committed table-page changes for an existing schema. DDL, DCL, VACUUM, migration or WAL rewind requires a new base archive.
  • One controller owns an archive directory; there is no multi-writer archive coordination.
  • Live archive generations retain complete WAL prefixes; rotate to a new base archive periodically to bound disk usage.

These are explicit scope limits, not silent fallbacks.

The deterministic storage failpoint validates MiniSQL's I/O-error handling but does not reproduce every filesystem or controller behavior on a genuinely full volume. Deployment-specific ENOSPC qualification must use a disposable, quota-limited volume that cannot exhaust the host operating system.