root 56844c3f31 embed cache — LRU at /v1/embed for repeat-query elimination
Adds CachedProvider wrapping the embedding Provider with a thread-safe
LRU keyed on (effective_model, sha256(text)) → []float32. Repeat
queries return the stored vector without round-tripping to Ollama.

Why this matters: the staffing 500K test (memory project_golang_lakehouse)
documented that the staffing co-pilot replays many of the same query
texts ("forklift driver IL", "welder Chicago", "warehouse safety", etc).
Each repeat paid the ~50ms Ollama round-trip. Cached repeats now serve
in <1µs (LRU lookup + sha256 of input).

Memory budget: ~3 KiB per entry at d=768. Default 10K entries ≈ 30 MiB.
Configurable via [embedd].cache_size; 0 disables (pass-through mode).

Per-text caching, not per-batch — a batch with mixed hits/misses only
fetches the misses upstream, then merges the result preserving caller
input order. Three-text batch with one miss = one upstream call for
that one text instead of three.

Implementation:
  internal/embed/cached.go (NEW, 150 LoC)
    CachedProvider implements Provider; uses hashicorp/golang-lru/v2.
    Key shape: "<model>:<sha256-hex>". Empty model resolves to
    defaultModel (request-derived) for the key — NOT res.Model
    (upstream-derived), so future requests with same input shape
    hit the same key. Caught by TestCachedProvider_EmptyModelResolvesToDefault.
    Atomic hit/miss counters + Stats() + HitRate() + Len().

  internal/embed/cached_test.go (NEW, 12 test funcs)
    Pass-through-when-zero, hit-on-repeat, mixed-batch only fetches
    misses, model-key isolation, empty-model resolves to default,
    LRU eviction at cap, error propagation, all-hits synthesized
    without upstream call, hit-rate accumulation, empty-texts
    rejected, concurrent-safe (50 goroutines × 100 calls), key
    stability + distinctness.

  internal/shared/config.go
    EmbeddConfig.CacheSize (toml: cache_size). Default 10000.

  cmd/embedd/main.go
    Wraps Ollama Provider with CachedProvider on startup. Adds
    /embed/stats endpoint exposing hits / misses / hit_rate / size.
    Operators check the rate to confirm the cache is working
    (high rate = good) or sized wrong (low rate + many misses on a
    workload that should have repeats).

  cmd/embedd/main_test.go
    Stats endpoint tests — disabled mode shape, enabled mode tracks
    hits + misses across repeat calls.

One real bug caught by my own test:
  Initial implementation cached under res.Model (upstream-resolved)
  rather than effectiveModel (request-resolved). A request with
  model="" caching under "test-model" (Ollama's default), then a
  request with model="the-default" (our config default) missing
  the cache. Fix: always use the request-derived effectiveModel
  for keys; that's the predictable side. Locked by
  TestCachedProvider_EmptyModelResolvesToDefault.

Verified:
  go test -count=1 ./internal/embed/  — all 12 cached tests + 6 ollama tests green
  go test -count=1 ./cmd/embedd/      — stats endpoint tests green
  just verify                          — vet + test + 9 smokes 33s

Production benefit:
  ~50ms Ollama round-trip → <1µs cache lookup for cached entries.
  At 10K-entry default + ~30% repeat rate (typical staffing co-pilot
  workload), saves several seconds per staffer-query session.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-04-29 06:54:30 -05:00

206 lines
7.3 KiB
Go
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

// Package shared also provides the TOML config loader. Per ADR
// equivalent of Rust ADR-006 (TOML config over env vars), every
// service reads `lakehouse.toml` with sane defaults and env
// overrides. Config is hot-reload-unaware in G0; reload-on-SIGHUP
// is a G1+ concern.
package shared
import (
"errors"
"fmt"
"io/fs"
"log/slog"
"os"
"github.com/pelletier/go-toml/v2"
)
// Config is the unified Lakehouse config. Each service reads only
// the section it cares about, but they all share the same file so
// operators have one place to look.
type Config struct {
Gateway GatewayConfig `toml:"gateway"`
Storaged ServiceConfig `toml:"storaged"`
Catalogd CatalogConfig `toml:"catalogd"`
Ingestd IngestConfig `toml:"ingestd"`
Queryd QuerydConfig `toml:"queryd"`
Vectord VectordConfig `toml:"vectord"`
Embedd EmbeddConfig `toml:"embedd"`
S3 S3Config `toml:"s3"`
Log LogConfig `toml:"log"`
}
// IngestConfig adds ingestd-specific knobs. ingestd needs to PUT
// parquet to storaged AND register manifests with catalogd, so it
// holds two upstream URLs in addition to its own bind.
//
// MaxIngestBytes caps the multipart body size. CSVs are typically
// 4-6× larger than the resulting Snappy-compressed Parquet, so 256
// MiB CSV → ~50 MiB Parquet — well under storaged's 256 MiB PUT
// cap. Real-scale validation (2026-04-29) showed 500K workers ×
// 18 cols = 344 MiB CSV → 71 MiB Parquet; bumping this knob to
// 512 MiB is the documented path for that workload.
type IngestConfig struct {
Bind string `toml:"bind"`
StoragedURL string `toml:"storaged_url"`
CatalogdURL string `toml:"catalogd_url"`
MaxIngestBytes int64 `toml:"max_ingest_bytes"`
}
// GatewayConfig adds the upstream URLs the reverse proxy fronts.
// Each route family (/v1/storage, /v1/catalog, /v1/ingest, /v1/sql,
// /v1/vectors, /v1/embed) has its own upstream so we can scale
// services independently or move them to different boxes without
// touching gateway code.
type GatewayConfig struct {
Bind string `toml:"bind"`
StoragedURL string `toml:"storaged_url"`
CatalogdURL string `toml:"catalogd_url"`
IngestdURL string `toml:"ingestd_url"`
QuerydURL string `toml:"queryd_url"`
VectordURL string `toml:"vectord_url"`
EmbeddURL string `toml:"embedd_url"`
}
// EmbeddConfig drives the embed service. ProviderURL points at the
// embedding backend (Ollama in G2, possibly OpenAI/Voyage in G3+).
// DefaultModel is what gets used when callers don't specify a
// model in their request body. CacheSize is the LRU cache cap on
// (model, sha256(text)) → vector lookups; 0 disables caching.
// Default 10000 entries ≈ 30 MiB at d=768.
type EmbeddConfig struct {
Bind string `toml:"bind"`
ProviderURL string `toml:"provider_url"`
DefaultModel string `toml:"default_model"`
CacheSize int `toml:"cache_size"`
}
// VectordConfig adds vectord-specific knobs. StoragedURL is
// optional — empty string disables persistence, useful for ephemeral
// dev or test runs. When set, indexes Save after every state change
// and Load on startup.
type VectordConfig struct {
Bind string `toml:"bind"`
StoragedURL string `toml:"storaged_url"`
}
// QuerydConfig adds queryd-specific knobs. queryd talks DuckDB
// directly to MinIO via DuckDB's httpfs extension (so no storaged
// URL needed), and reads the catalog over HTTP for view registration.
// SecretsPath defaults to /etc/lakehouse/secrets-go.toml — the same
// file storaged uses, since both services need the S3 credentials.
type QuerydConfig struct {
Bind string `toml:"bind"`
CatalogdURL string `toml:"catalogd_url"`
SecretsPath string `toml:"secrets_path"`
RefreshEvery string `toml:"refresh_every"` // duration string, e.g. "30s"
}
// CatalogConfig adds catalogd-specific knobs on top of the standard
// bind. StoragedURL points at the storaged service for manifest
// persistence; G0 defaults to the localhost bind.
type CatalogConfig struct {
Bind string `toml:"bind"`
StoragedURL string `toml:"storaged_url"`
}
// ServiceConfig is the per-binary bind config. Default Bind ""
// means "use the service's hardcoded G0 default" — see DefaultConfig.
type ServiceConfig struct {
Bind string `toml:"bind"`
}
// S3Config holds S3-compatible storage settings. Endpoint blank →
// AWS default. Bucket "" → "lakehouse-primary".
type S3Config struct {
Endpoint string `toml:"endpoint"`
Region string `toml:"region"`
Bucket string `toml:"bucket"`
AccessKeyID string `toml:"access_key_id"`
SecretAccessKey string `toml:"secret_access_key"`
UsePathStyle bool `toml:"use_path_style"`
}
// LogConfig — slog level for now; structured fields land G1+.
type LogConfig struct {
Level string `toml:"level"`
}
// DefaultConfig returns the G0 dev defaults. Ports are shifted to
// 3110+ to coexist with the live Rust lakehouse on 3100/3201-3204
// during the migration. G5 cutover flips gateway back to 3100.
func DefaultConfig() Config {
return Config{
Gateway: GatewayConfig{
Bind: "127.0.0.1:3110",
StoragedURL: "http://127.0.0.1:3211",
CatalogdURL: "http://127.0.0.1:3212",
IngestdURL: "http://127.0.0.1:3213",
QuerydURL: "http://127.0.0.1:3214",
VectordURL: "http://127.0.0.1:3215",
EmbeddURL: "http://127.0.0.1:3216",
},
Storaged: ServiceConfig{Bind: "127.0.0.1:3211"},
Catalogd: CatalogConfig{Bind: "127.0.0.1:3212", StoragedURL: "http://127.0.0.1:3211"},
Ingestd: IngestConfig{
Bind: "127.0.0.1:3213",
StoragedURL: "http://127.0.0.1:3211",
CatalogdURL: "http://127.0.0.1:3212",
MaxIngestBytes: 256 << 20, // 256 MiB; bump per deployment via lakehouse.toml
},
Vectord: VectordConfig{
Bind: "127.0.0.1:3215",
StoragedURL: "http://127.0.0.1:3211",
},
Embedd: EmbeddConfig{
Bind: "127.0.0.1:3216",
ProviderURL: "http://localhost:11434", // local Ollama
DefaultModel: "nomic-embed-text",
CacheSize: 10_000, // ~30 MiB at d=768; set to 0 to disable
},
Queryd: QuerydConfig{
Bind: "127.0.0.1:3214",
CatalogdURL: "http://127.0.0.1:3212",
SecretsPath: "/etc/lakehouse/secrets-go.toml",
RefreshEvery: "30s",
},
S3: S3Config{
Endpoint: "http://localhost:9000",
Region: "us-east-1",
Bucket: "lakehouse-primary",
UsePathStyle: true,
},
Log: LogConfig{Level: "info"},
}
}
// LoadConfig reads `lakehouse.toml` from path; if path is empty or
// the file doesn't exist, returns DefaultConfig. Any decode error is
// fatal (we don't want a misconfigured service silently falling back
// to defaults — that's the kind of bug you find at 2am).
//
// Per Opus + Qwen WARN #3: when path WAS given but the file is
// missing, log a warning so silent default-fallback doesn't hide
// misconfiguration. Empty path is fine (caller didn't ask for a
// file); non-empty + missing is suspicious.
func LoadConfig(path string) (Config, error) {
cfg := DefaultConfig()
if path == "" {
return cfg, nil
}
b, err := os.ReadFile(path)
if errors.Is(err, fs.ErrNotExist) {
slog.Warn("config file not found, using defaults",
"path", path,
"hint", "create the file or pass -config /path/to/lakehouse.toml")
return cfg, nil
}
if err != nil {
return cfg, fmt.Errorf("read config: %w", err)
}
if err := toml.Unmarshal(b, &cfg); err != nil {
return cfg, fmt.Errorf("parse config: %w", err)
}
return cfg, nil
}