Skip to content

BENCHMARKS.md

Latest commit

 

History

History
134 lines (93 loc) · 4.09 KB

BENCHMARKS.md

File metadata and controls

134 lines (93 loc) · 4.09 KB

Benchmarks

Performance evolution of AnvilDB across versions.

Run benchmarks yourself:

php benchmarks/benchmark.php 10000

v0.4.0 — New operators, aggregations, range indexes, CSV

2026-03-16 | PHP 8.4.18 | Rust stable | Linux x86_64 | 10,000 records

CRUD

Operation Time Throughput
Bulk insert (10x1000) 214ms ~47k docs/s
Read all (10k docs) 22ms ~448k docs/s

Queries

Operation Time Results
Filter (= admin) 4.6ms 2,500
Filter + sort + limit(100) 3.6ms 100
whereBetween(age, 30, 50) 5.1ms 2,625
whereIn(role, 2 values) 9.2ms 5,000
whereRegex(name, pattern) 71.6ms 100
Count with filter 0.4ms 6,666

Aggregations

Operation Time
sum + avg + min + max 4.1ms
group_by(role) + count + avg 7.9ms

Indexes

Operation Time
Create range index 5.4ms
whereBetween with range index 4.6ms

All operations include compression, atomic writes, schema validation, and index enforcement.

Changes

  • New query operators: between, in, not_in
  • Aggregations: sum, avg, min, max, count with optional group_by
  • Range indexes (BTreeMap) for ordered lookups on >, <, >=, <=, between
  • CSV export/import
  • Updates and deletes now buffered (mark dirty + background flush)

v0.3.0 — Compression, encryption, joins, lazy loading

2026-03-16 | PHP 8.4.18 | Rust stable | Linux x86_64 | 10,000 records

Operation Time Throughput vs v0.2.0
Bulk insert (10x1000) 199ms ~50k docs/s ~same
Read all (10k docs) 23ms ~441k docs/s 5.4x faster
Filter (= admin) 4.3ms — 7.9x faster
Filter + sort + limit 3.4ms — 4.6x faster
Count with filter 0.2ms — 9x faster

All operations include compression, atomic writes, schema validation, and index enforcement.

Changes

  • Deflate compression on all collection and index files (.anvil format)
  • Optional AES-256-GCM encryption at rest
  • INNER and LEFT joins across collections (hash join)
  • Lazy loading — collections loaded on first access, not at startup
  • Write buffer simplified to dirty-set tracking with full rewrite on flush
  • FFI functions moved from lib.rs to ffi.rs

Impact

  • Read/query performance dramatically improved — lazy loading + compressed format means less I/O
  • Insert throughput maintained despite switching from append to full rewrite (compression reduces file size)

v0.2.0 — NDJSON append-only storage

2026-03-15 | PHP 8.4.18 | Rust stable | Linux x86_64 | 10,000 records

Operation Time Throughput vs v0.1.0
Bulk insert (10x1000) 222ms ~45k docs/s 4.9x faster
Read all (10k docs) 125ms ~80k docs/s ~same
Filter (= admin) 34ms — ~same
Filter + sort + limit 16ms — ~same
Count with filter 1.8ms — ~same

All operations include atomic writes, compression, schema validation, and index enforcement.

Changes

  • Storage migrated from JSON array to NDJSON (one JSON object per line)
  • Inserts use append mode — O(1) per document instead of O(n) full rewrite
  • Automatic migration from legacy format on first open

Impact

  • Bulk insert 4.9x faster (1085ms → 222ms)
  • Read/query performance unchanged (expected — reads still parse full file)

v0.1.0 — JSON array storage + atomic read-modify-write

2026-03-15 | PHP 8.4.18 | Rust stable | Linux x86_64 | 10,000 records

Operation Time Throughput
Bulk insert (10x1000) 1085ms ~9k docs/s
Read all (10k docs) 123ms ~81k docs/s
Filter (= admin) 33ms —
Filter + sort + limit 14ms —
Count with filter 1.5ms —

Notes

  • Every write operation uses atomic read-modify-write (exclusive file lock + temp file + rename), re-reading and re-writing the entire JSON array per batch
  • This guarantees multi-process safety at the cost of write throughput
  • Write performance was the main bottleneck, addressed in v0.2.0 with NDJSON