Files
OmniRoute/src/lib/db/readCache.ts
Diego Rodrigues de Sa e Souza d776a42324 fix(api): invalidate /v1/models + specialty catalogs on DB writes; fix duplicated headers and 500 replayed as 200 (#6408)
The #6408 request-shape-keyed TTL cache around getUnifiedModelsResponse was not
keyed by DB/settings state, so a write followed by a read within the ~1.5s TTL
replayed the pre-write catalog (dropping newly-eligible models like codex/gpt-5.5
and every specialty catalog routed through it since #6303). Fold a
modelCatalogCacheVersion (bumped by invalidateDbCache, already called on every
settings/connection/combo/pricing write) into the cache so a state change forces
an immediate miss; merge response headers through a real Headers instance
(fixes X-Request-Id duplication); carry and replay status end-to-end (a mid-build
500 was replayed as 200). Tests call the existing __resetCatalogBuilderRunsForTest
hook in setup, matching v1-models-concurrent-6408.
2026-07-07 07:41:20 -03:00

203 lines
7.2 KiB
TypeScript

/**
* DB Read Cache — In-memory TTL cache for hot read paths.
*
* SQLite reads are already fast since better-sqlite3 is synchronous and
* memory-mapped. However, some functions (getSettings, getPricing,
* getProviderConnections) are called on every request by multiple callers.
* A short TTL cache (5s) eliminates redundant I/O without staling data for
* long enough to matter (settings changes are applied within one cache cycle).
*
* Usage:
* import { dbCache } from '@/lib/db/readCache';
* const settings = await dbCache.getSettings();
*/
type CacheEntry<T> = {
value: T;
expiresAt: number;
};
class TTLCache<T> {
private cache = new Map<string, CacheEntry<T>>();
private readonly ttlMs: number;
constructor(ttlMs: number) {
this.ttlMs = ttlMs;
}
get(key: string): T | undefined {
const entry = this.cache.get(key);
if (!entry) return undefined;
if (Date.now() > entry.expiresAt) {
this.cache.delete(key);
return undefined;
}
return entry.value;
}
set(key: string, value: T): void {
this.cache.set(key, { value, expiresAt: Date.now() + this.ttlMs });
}
invalidate(key?: string): void {
if (key) {
this.cache.delete(key);
} else {
this.cache.clear();
}
}
}
// Cache with 5s TTL — short enough to pick up dashboard changes quickly,
// long enough to serve burst request bursts without hammering SQLite.
const SETTINGS_TTL_MS = 5_000;
const PRICING_TTL_MS = 30_000;
const CONNECTIONS_TTL_MS = 5_000;
const settingsCache = new TTLCache<Record<string, unknown>>(SETTINGS_TTL_MS);
const pricingCache = new TTLCache<Record<string, unknown>>(PRICING_TTL_MS);
const connectionsCache = new TTLCache<unknown[]>(CONNECTIONS_TTL_MS);
/**
* Cached wrapper for getSettings.
* Invalidated on every updateSettings() call.
*/
export async function getCachedSettings(): Promise<Record<string, unknown>> {
const cached = settingsCache.get("settings");
if (cached) return cached;
const { getSettings } = await import("@/lib/db/settings");
const value = await getSettings();
settingsCache.set("settings", value);
return value;
}
/**
* Cached wrapper for getPricing.
* Longer TTL since pricing rarely changes mid-session.
*/
export async function getCachedPricing(): Promise<Record<string, unknown>> {
const cached = pricingCache.get("pricing");
if (cached) return cached as Record<string, unknown>;
const { getPricing } = await import("@/lib/db/settings");
const value = await getPricing();
pricingCache.set("pricing", value);
return value;
}
/**
* Cached wrapper for getProviderConnections.
* Used in request hot-paths (usageStats, callLogs, usageHistory).
*/
export async function getCachedProviderConnections(
filter?: Record<string, unknown>
): Promise<unknown[]> {
// Only cache the unfiltered "all connections" query (most common)
if (filter && Object.keys(filter).length > 0) {
const { getProviderConnections } = await import("@/lib/db/providers");
return getProviderConnections(filter);
}
const cached = connectionsCache.get("all");
if (cached) return cached;
const { getProviderConnections } = await import("@/lib/db/providers");
const value = await getProviderConnections();
connectionsCache.set("all", value);
return value;
}
// ──────────────── LKGP Cache Wrappers ────────────────
interface LKGPRecordCache {
provider: string;
connectionId?: string;
}
const lkgpCache = new TTLCache<LKGPRecordCache | null>(SETTINGS_TTL_MS);
export async function getCachedLKGP(
comboName: string,
modelId: string
): Promise<LKGPRecordCache | null> {
const cacheKey = `lkgp:${comboName}:${modelId}`;
const cached = lkgpCache.get(cacheKey);
if (cached !== undefined) return cached;
const { getLKGP } = await import("@/lib/db/settings");
const value = await getLKGP(comboName, modelId);
lkgpCache.set(cacheKey, value);
return value;
}
export async function setCachedLKGP(
comboName: string,
modelId: string,
providerId: string,
connectionId?: string
): Promise<void> {
const { setLKGP } = await import("@/lib/db/settings");
await setLKGP(comboName, modelId, providerId, connectionId);
lkgpCache.invalidate(`lkgp:${comboName}:${modelId}`);
}
// ──────────────── Combo Cache Invalidation Signal ────────────────
//
// The nested-combo expansion caches live in request handlers
// (`src/sse/handlers/chat.ts` getCombosCachedForChat and
// `open-sse/handlers/chatCore.ts` getCombosCached), each with a 10s TTL. A db
// module must NOT import a request handler (that would create an import cycle),
// so instead those caches consult this monotonically-incrementing version.
// Combo writes call `invalidateDbCache("combos")`, which bumps the version;
// the handlers compare the version they were populated at against the current
// one and treat a mismatch as a cache miss — so combo edits take effect
// immediately instead of after the 10s window (#3147).
let combosCacheVersion = 0;
/**
* Current combo-cache version. Cache layers snapshot this when they populate
* and re-read it on every access; a change means the underlying combos were
* written and the cached expansion must be refreshed.
*/
export function getCombosCacheVersion(): number {
return combosCacheVersion;
}
// ──────────────── Model Catalog Cache Invalidation Signal ────────────────
//
// #6408 added a request-shape-keyed (prefix/isCodex/apiKey) TTL cache around the
// unified /v1/models builder (src/app/api/v1/models/catalog.ts) to coalesce
// concurrent/bursty GETs. That cache key does not vary with the underlying DB
// state the builder reads (connections, settings, combos), so a write followed by
// a read within the ~1.5s TTL replayed the pre-write response. Same import-cycle
// constraint as combosCacheVersion above (a db module must not import the route
// module) — catalog.ts instead compares this version on every access and drops its
// whole cache the moment it moves, so any write that calls invalidateDbCache() makes
// the next read miss immediately instead of waiting out the TTL.
let modelCatalogCacheVersion = 0;
/**
* Current model-catalog-cache version. `getUnifiedModelsResponse()` folds this
* into its response cache key; a change means settings/connections/combos were
* written since the cache was populated and the cached body is stale.
*/
export function getModelCatalogCacheVersion(): number {
return modelCatalogCacheVersion;
}
/**
* Invalidate all caches (call after writes to any of: settings, pricing,
* connections, combos).
*/
export function invalidateDbCache(scope?: "settings" | "pricing" | "connections" | "combos"): void {
if (!scope || scope === "settings") settingsCache.invalidate();
if (!scope || scope === "pricing") pricingCache.invalidate();
if (!scope || scope === "connections") connectionsCache.invalidate();
if (!scope || scope === "combos") combosCacheVersion++;
// Settings/connections/combos all feed the unified model catalog builder
// (blockedProviders + hidePaidModels, provider connections + excludedModels,
// combo definitions, respectively) — pricing does too, via isFreeModel().
modelCatalogCacheVersion++;
}