Skip to content
DocspackagesDocumentation

@lunora/hyperdrive

Bring-your-own Postgres/MySQL via Cloudflare Hyperdrive — an action-only ctx.sql, or a reactive .global() backend.

PackagesHyperdrive

@lunora/hyperdrive lets an action read and write an existing Postgres/MySQL database through Cloudflare Hyperdrive: pooled, cached connections from the edge. It surfaces the binding's connection string and a driver-agnostic ctx.sql client.

Integrate an existing database; don't replace the Lunora data layer. Hyperdrive points at a database Lunora has no visibility into, so two invariants the rest of Lunora relies on do not hold for it:

  1. Non-deterministic. A SQL query over the network is an external, mutable read, exactly like fetch. It is therefore forbidden in query/mutation and available only on ActionCtx. The hyperdrive_outside_action advisor lint flags any ctx.sql reached from a query or mutation.
  2. Non-reactive. Live queries track writes to the Durable Object (DO)'s SQLite / D1. An UPDATE issued over Hyperdrive produces no Lunora change event, so subscriptions will not re-run when external rows change.

Hyperdrive is the right tool for "read/write my legacy Postgres from an action," and the wrong tool for "make my Postgres reactive." If you want external data to be reactive, write a projection of it into a defineSchema DO/D1 table (see Making external data reactive).

Install

pnpm add @lunora/hyperdrive

No driver is bundled: postgres, pg, and mysql2 are heavy and the choice is yours. They are declared as optional peer dependencies; install the one you use:

pnpm add postgres        # postgres.js  → fromPostgresJs
# or
pnpm add pg              # node-postgres → fromNodePg
# or
pnpm add mysql2          # mysql2        → fromMysql2

Set up the binding

Importing @lunora/hyperdrive (or reading ctx.sql) is enough: lunora dev and lunora deploy add a HYPERDRIVE binding to wrangler.jsonc for you, with a placeholder id.

  • In dev, ctx.sql runs against an embedded Postgres (PGlite), with its data in .wrangler/state/lunora-pglite, so there is nothing to run. It starts under lunora dev, vite dev and vite preview. It needs two optional packages, and lunora dev says so when they are missing:

    pnpm add -D @electric-sql/pglite @electric-sql/pglite-socket

    To use your own database instead, set CLOUDFLARE_HYPERDRIVE_LOCAL_CONNECTION_STRING_HYPERDRIVE (the last part is the binding name) or localConnectionString on the binding. Neither is overridden. The embedded database speaks PostgreSQL: a project whose only SQL driver is mysql2 gets a warning and must set that variable to a local MySQL of its own. A plain wrangler dev that Lunora does not start has no embedded database either: set the variable (or localConnectionString) there too.

  • On deploy, set the origin's connection string and a token that may edit Hyperdrive (CLOUDFLARE_API_TOKEN with Hyperdrive: Edit, plus CLOUDFLARE_ACCOUNT_ID):

    LUNORA_HYPERDRIVE_URL="postgres://user:pass@host:5432/db" lunora deploy # gitleaks:allow -- placeholder, not a real secret

    The variable is LUNORA_<BINDING>_URL. Lunora creates a config named after the deployed Worker (so --env staging gets its own) and the binding, or reuses one with that name that reaches the same database, and writes its id into wrangler.jsonc. A same-named config for another database stops the deploy. The connection string is only sent to Cloudflare: it is never written to a file or logged. Once the id is real, later deploys need neither variable.

Prefer to manage it yourself? Run wrangler hyperdrive create my-db --connection-string=... and put the id in the binding; Lunora then leaves it alone. Lunora validates the binding (it errors when binding is missing).

Hyperdrive caches read results for up to a minute by default, and a write does not invalidate them, so a ctx.sql read can miss a row your own action just wrote. For reads that must be fresh, follow Cloudflare's advice and add a second config created with --caching-disabled, bound under another name. A .global() backend always needs caching off: see below.

The canonical recipe

ctx.sql is wired once, at the app level, not inside a handler: the generated ctx.sql is a readonly property (assigning to it is a TS2540). Codegen cannot build the client for you — createHyperdrive returns connection info, and turning that into a SqlClient needs the driver you chose — so it emits a sql config thunk instead. Wire it on the app builder:

src/server/index.ts
import type { HyperdriveLike } from "@lunora/hyperdrive";
import { createHyperdrive, fromPostgresJs } from "@lunora/hyperdrive";

import { defineApp } from "../../lunora/_generated/app";

const app = defineApp<Env>()
    .shard((env) => env.SHARD)
    // Called once per shard construction, not per request.
    .hyperdrive((env) =>
        fromPostgresJs(async () => {
            const { default: postgres } = await import("postgres");

            return postgres(createHyperdrive(env.HYPERDRIVE as HyperdriveLike).connectionString, { fetch_types: false, max: 5 });
        }),
    )
    .build();

export const ShardDO = app.ShardDO;

Hand-composing the worker instead? The same thunk is createShardDO({ sql }):

export const ShardDO = createShardDO({
    sql: (env) =>
        fromPostgresJs(async () => {
            const { default: postgres } = await import("postgres");

            return postgres(createHyperdrive(env.HYPERDRIVE as HyperdriveLike).connectionString, { fetch_types: false, max: 5 });
        }),
});

The adapter takes a loader, not a driver: it runs on the first query and its client is reused, so the driver stays out of the Worker's startup graph. A static import postgres from "postgres" in the entry is evaluated at startup and defeats this. A failed load rejects that query; the next one retries. For pg, the loader may await client.connect() before returning the client.

Without the thunk ctx.sql is a stub whose every method throws a message pointing back at this wiring.

Handlers then just read it. Only inside an action:

import { action, v } from "@/lunora/_generated/server";

export const importLegacyOrders = action.input({ orgId: v.string() }).action(async ({ ctx, args: { orgId } }) => {
    // Read from external Postgres ($1, $2, … placeholders).
    const orders = await ctx.sql.query<{ id: string; total: number }>("select id, total from orders where org = $1", [orgId]);

    return orders;
});

When the codegen detects ctx.sql usage it adds sql: SqlClient to ActionCtx only (never QueryCtx or MutationCtx), with a JSDoc restating the determinism/realtime caveat.

Drivers & placeholders

DriverAdapterPlaceholders
postgres (postgres.js)fromPostgresJs$1, $2, …
pg (node-postgres)fromNodePg$1, $2, …
mysql2/promisefromMysql2?

The package never rewrites SQL; use your driver's native positional syntax.

Build postgres.js with { fetch_types: false, max: 5 }, Cloudflare's recommended Hyperdrive setup: max stays under the Workers limit on concurrent connections, and fetch_types: false saves a round trip on connect. fromPostgresJs runs every statement as a named prepared statement (unsafe's own default is unnamed, which Hyperdrive's transaction pooling does not reliably support). Prepared reads are cacheable, so a ctx.sql read through postgres.js can be served from Hyperdrive's query cache like one through pg.

fetch_types: false leaves postgres.js without array types, so a JS array bound to = ANY($1) goes over as a malformed literal. Bind a list as JSON text instead (IN (SELECT jsonb_array_elements_text($1::text::jsonb)) with JSON.stringify(list)), or drop the option if you rely on array params. Cast a JSON parameter $1::text::jsonb, not bare $1::jsonb: postgres.js types the parameter from the server and would JSON-encode your already-serialised string a second time.

fromMysql2 (and buildMysqlExec) call mysql2's query, not execute: Hyperdrive does not support MySQL protocol-level prepared statements, so the ? values are escaped client-side and the statement goes over as text. An array or object value binds as its JSON text (mysql2 alone would expand it into SQL), and an undefined value throws. Strings are quoted with a doubled '' for a quote and \\ for a backslash, so the server must not run with NO_BACKSLASH_ESCAPES in its sql_mode (MySQL's default doesn't include it): under that mode a backslash would be stored twice. A quote can't end the literal under either mode. A ? inside a quoted string or a comment is left alone, but ?? is an identifier placeholder (the value is quoted with backticks). Create the mysql2 connection with disableEval: true, which Workers require.

Telemetry

With an observability sink configured, ctx.sql is instrumented automatically at the same level as ctx.db — the sink's instrumentDatabase field (see Observability: database telemetry):

  • "summary" (default) folds sql.calls, sql.duration_ms, sql.errors and a per-operation count (sql.op.SELECT, sql.op.INSERT, …) onto the dispatch's root span. These are separate from ctx.db's db.* counters, so time spent in your external database stays distinguishable from the shard's own SQLite.
  • "spans" adds one CLIENT span per query (sql.SELECT, …), capped at 100 per dispatch (sql.spans_truncated past that), carrying db.system.name (postgresql for fromPostgresJs / fromNodePg, mysql for fromMysql2, other_sql for a hand-built client), db.operation.name and db.response.returned_rows.
  • "off" hands the action your client untouched.

The operation is the statement's leading keyword only. Statement text and parameters are never recorded, there is no db.collection.name, and a failed query records its error type rather than the driver's message (outside development), since Postgres and MySQL errors quote row values. SqlClient has a single query method, so there are no transaction or batch spans — a BEGIN/COMMIT you issue yourself is counted under its own keyword.

Queries you run through your own driver or ORM instance outside ctx.sql are not seen. For a Drizzle instance on Hyperdrive, cloudflare-drizzle-tracing's instrumentDrizzle() adds Cloudflare-native drizzle.execute / transaction / batch spans — be aware it records query text and parameters.

Making external data reactive

Lunora cannot observe external writes, so a subscription over a defineSchema table won't re-fire when Postgres changes. To make external data reactive, project it into a DO/D1 table inside the same action. That write is on the change-feed, so live queries reading the projection re-run:

import { api } from "@/lunora/_generated/api";
import { action, v } from "@/lunora/_generated/server";

export const syncOrder = action.input({ id: v.string() }).action(async ({ ctx, args: { id } }) => {
    const [row] = await ctx.sql.query<{ id: string; total: number }>("select id, total from orders where id = $1", [id]);

    // This write is tracked — a `query` over `orders` re-runs for subscribers.
    await ctx.runMutation(api.orders.upsert, { id: row.id, total: row.total });
});

Per-agent shape ingest (multitenant Postgres → per-tenant DOs)

A common shape of the pattern above: a multitenant Postgres is the source of truth, and you want each tenant (or agent) to work against only its own slice, materialized into its own sharded Durable Object, with clients riding the same live slice. .shardBy() gives each tenant a private DO + SQLite; pullSourceRows (read side) and materializeExternalRows (write side) bridge the Postgres slice into it; defineShape carries it to clients with no extra wiring.

// lunora/schema.ts — one DO per tenant, plus a shape clients subscribe to
export default defineSchema({
    documents: defineTable({ title: v.string(), body: v.string(), orgId: v.string() }).shardBy("orgId").externallyManaged(), // rows are written by the ingest bridge, not user mutations
});

export const tenantDocs = defineShape({ table: "documents", where: () => ({}) });
// The action pulls THIS tenant's slice (the shard key binds into the WHERE — the
// tenant-isolation boundary), then a mutation materializes it.
import { pullSourceRows } from "@lunora/hyperdrive";

import { api } from "@/lunora/_generated/api";
import { action, mutation, v } from "@/lunora/_generated/server";
import { materializeExternalRows } from "@lunora/shard-engine";

export const refreshTenant = action.input({ orgId: v.string() }).action(async ({ ctx, args: { orgId } }) => {
    const docs = await pullSourceRows(ctx.sql, {
        query: 'select id, title, body, org_id as "orgId" from documents where org_id = $1',
        params: [orgId], // ← tenant scope. NEVER omit this on a sharded source.
        map: (row) => ({ body: row.body, orgId: row.orgId, title: row.title }),
    });

    // One mutation, so the whole slice materializes atomically: a failure part
    // way through leaves the tenant's DO on its previous snapshot, not half-updated.
    await ctx.runMutation(api.documents.ingest, { orgId, docs });
});

export const ingest = mutation.input({ orgId: v.string(), docs: v.array(v.any()) }).mutation(async ({ ctx, args: { docs } }) => {
    // Empty baseline = upsert-only (inserts + updates, no deletes). The materialized
    // rows append to the CDC log, so `tenantDocs` subscribers are poked live.
    await materializeExternalRows(ctx.db, docs, new Map(), { table: "documents" });
});

Clients consume the same slice with the shape they'd use over any table:

const docs = useShape(api.shapes.tenantDocs, {}); // RLS-filtered, live, per-tenant

Refresh on a schedule with @lunora/scheduler (ctx.scheduler.runAfter(...) / a cron) so each tenant's slice stays current.

Tenant scoping is the correctness boundary. Per-shard SQLite isolation only controls where rows land, not what the query pulls. The shard key MUST bind into the source WHERE (as a parameter), or every tenant's DO would replicate the whole table. Deletes: the empty-baseline call above is upsert-only; to propagate upstream deletes, pass the table's current membership as the baseline so materializeExternalRows can diff it.

Or skip the boilerplate: the declarative .source() modifier

The manual bridge above is the escape hatch. For the common case, declare the source on the table and Lunora runs the whole loop on the DO's alarm: full-pull diff, tenant scoping, and a poll cadence, with no action, mutation, or cron to write:

// lunora/schema.ts
export default defineSchema({
    documents: defineTable({ title: v.string(), body: v.string(), orgId: v.string() })
        .shardBy("orgId")
        .source({
            binding: "HYPERDRIVE_DOCS",
            query: 'select id, title, body, org_id as "orgId" from documents where org_id = $1',
            tenantBy: (shardKey) => [shardKey], // mandatory under .shardBy() — the tenant boundary
        }),
});

export const tenantDocs = defineShape({ table: "documents", where: () => ({}) });

You provide the driver once, on the app builder: one resolver for every sourced binding (build the SqlClient exactly as above). Lunora memoizes it per binding and polls each tenant's slice on the alarm. It is required — without it every tick records no sourceClient resolved for binding "…" and the table stays empty.

// worker.ts
export default defineApp<Env>()
    .shard((env) => env.SHARD)
    .sourceClient((env, binding) =>
        fromPostgresJs(async () => {
            const { default: postgres } = await import("postgres");

            return postgres((env[binding] as { connectionString: string }).connectionString, { fetch_types: false, max: 5 });
        }),
    )
    .build();

.sourceClient() appears on the builder only when your schema declares a .source(...) table. Composing the shard DO by hand instead, pass the same resolver as createShardDO({ sourceClient }).

Tenant scoping is enforced two ways: defineSchema throws at load if a sourced .shardBy() table omits tenantBy (the runtime fail-safe), and the external_source_unscoped advisor lint flags it earlier, at build time, in your terminal + Studio. defineSchema likewise rejects combining .source() with .global() (contradictory tiers). Clients consume the slice with useShape(api.shapes.tenantDocs), with no client change, since it is an ordinary table.

Refresh cadence. Omit refresh to poll on every DO alarm tick (the floor), or pass refresh: { everyMs } to throttle a large slice to at most one pull per interval. refresh: "manual" tells Lunora not to auto-poll; use it when you drive the refresh yourself (the manual pullSourceRows → materializeExternalRows bridge above, on your own schedule).

Delete detection: mode. The default mode: "full-pull" reads the whole tenant slice each tick and diffs it, so upstream deletes are observed for free, but it costs a full read per tick. For a large, low-churn slice, mode: "incremental" pulls only rows changed since a durable watermark:

documents: defineTable({ title: v.string(), body: v.string(), orgId: v.string(), updatedAt: v.number() })
    .shardBy("orgId")
    .source({
        binding: "HYPERDRIVE_DOCS",
        mode: "incremental",
        query: 'select id, title, body, org_id as "orgId", updated_at as "updatedAt" from documents where org_id = $1',
        // cursor.query pulls only rows past the watermark — tenantBy params bind first ($1), the watermark last ($2).
        cursor: {
            column: "updatedAt",
            query: 'select id, title, body, org_id as "orgId", updated_at as "updatedAt" from documents where org_id = $1 and updated_at >= $2 order by updated_at',
        },
        // Delete visibility — REQUIRED for incremental (pick one):
        reconcileEveryMs: 3_600_000, // periodic full-pull sweep GCs upstream deletes, OR:
        // softDeleteColumn: "deletedAt", // upstream tombstone column the cursor query returns
        tenantBy: (shardKey) => [shardKey],
    }),

How it runs: the first poll (and every reconcileEveryMs sweep) does a full-pull to seed membership + the watermark and GC deletes; every other tick binds the stored watermark as the cursor query's trailing param and upserts the returned slice. The watermark (the max cursor.column value seen) and the last-reconcile time persist per (table, shard) in the DO's reserved __lunora_source_cursor table, so they survive hibernation. Use >= in the cursor query (rows sharing a boundary timestamp re-pull idempotently rather than being skipped).

Because an incremental slice can't see a delete (an absent row means "unchanged", not "deleted"), incremental requires a delete-visibility path: reconcileEveryMs (a periodic full-pull sweep) or softDeleteColumn (an upstream tombstone column the cursor query returns, turned into a local delete; don't filter it out of the query). defineSchema throws, and the external_source_incremental_no_delete_path advisor lint fails the build, if an incremental source declares neither.

The cursor column must be strictly commit-monotonic: a value that only ever increases as rows become visible, never assigned below the current max after a later commit. A plain updated_at set from wall-clock time can violate this under clock skew or a long transaction (a row commits with a timestamp below a watermark already advanced past it), and an incremental slice would then miss it permanently. A reconcileEveryMs sweep is the backstop that re-establishes such rows; softDeleteColumn alone does not (it only catches deletes, not missed inserts), so prefer reconcileEveryMs unless your cursor is provably monotonic (e.g. a gapless sequence or a logical replication LSN).

Reactive .global() over Hyperdrive

The @lunora/hyperdrive/global subpath is the inverse trade-off: instead of an escape hatch onto a DB Lunora doesn't own, it makes a Postgres/MySQL database a reactive .global() storage backend, alongside D1. Lunora owns the schema: a .global() table gets a real column-per-field layout and every write routes through the shared store core, so live queries stay reactive with no extra wiring.

You build the writer inside the Durable Object that hosts the .global() store (the HYPERDRIVE binding is reachable there) and inject it as globalDb. Cache the driver on the DO instance and rebuild it lazily after hibernation.

import { fromPostgresJs } from "@lunora/hyperdrive";
import { createPostgresGlobalCtxDb } from "@lunora/hyperdrive/global";

// The loader runs on the first statement, so postgres.js is not evaluated at Worker startup.
const globalDb = createPostgresGlobalCtxDb(
    fromPostgresJs(async () => {
        const { default: postgres } = await import("postgres");

        return postgres(env.HYPERDRIVE.connectionString, { fetch_types: false, max: 5 });
    }),
    { schema },
);

A .global() write commits its row and companion rows in one transaction, so the driver must be able to open one: a postgres.js client with begin(), or a mysql2 pool (createPool, not a single createConnection). Anything else is refused with GLOBAL_TRANSACTION_UNSUPPORTED instead of committing statement by statement.

Turn Hyperdrive's query cache off

The Hyperdrive config behind a .global() store must have query caching disabled. Hyperdrive caches read results by default (max_age 60 s plus 15 s of stale_while_revalidate), and a write does not invalidate that cache. The store's own reads would then lag its writes: an optimistic-concurrency re-read returns the old version, so the conditional write misses and every retry rereads the same cached row (spurious ConflictErrors), and change-log polls and live-query re-runs serve rows up to 75 s old.

lunora deploy creates (or updates) its configs with caching off when the schema has .global({ backend: "hyperdrive" }) tables. For a binding whose id is already set, it doesn't change the config; with CLOUDFLARE_API_TOKEN and CLOUDFLARE_ACCOUNT_ID in the environment it warns when caching is still on. A config you made yourself needs it set once:

wrangler hyperdrive create my-db --connection-string="postgres://user:pass@host/db" --caching-disabled # gitleaks:allow -- placeholder
# or, for an existing config:
wrangler hyperdrive update <id> --caching-disabled

Local dev (localConnectionString, the embedded Postgres) has no cache, so it won't reproduce the problem. Node and celld have no Hyperdrive: their .global() store is SQLite (a local file on Node, a D1 binding on celld), so this does not apply there.

Tables provisioned by an earlier version

The store creates missing tables and adds missing columns on its own, but it does not rewrite a column that already exists. Two gaps can remain on a table provisioned before a release changed the column it would declare:

  • A field that accepts null but whose column is NOT NULL. Required v.any(), v.null(), v.literal(null) and unions with a nullable member used to be provisioned NOT NULL, so storing null there fails. Postgres drops the constraint itself at startup (a catalog-only change). MySQL does not: its only way to relax a column is MODIFY COLUMN, which restates the whole column and can rebuild the table. The store logs one warning per column with the exact statement to run instead, built from the column's current type and collation, for example:

    ALTER TABLE `notes` MODIFY COLUMN `note` longtext CHARACTER SET utf8mb4 COLLATE utf8mb4_0900_ai_ci NULL;

    Run it once per table in a maintenance window.

  • MySQL collation. Text columns are now declared utf8mb4_0900_bin, which compares byte for byte like SQLite and Postgres. A column created earlier keeps the server default (usually utf8mb4_0900_ai_ci, case- and accent-insensitive). Convert a table with ALTER TABLE <t> CONVERT TO CHARACTER SET utf8mb4 COLLATE utf8mb4_0900_bin. Until then, an in filter too long for one statement's placeholders compares under the column's own collation on utf8mb4_0900_ai_ci and _bin columns. On any other legacy collation (utf8mb4_general_ci, utf8mb4_unicode_ci) MySQL refuses such a filter with "Illegal mix of collations" instead of answering it — convert the table.

.searchIndex() works on a Hyperdrive-backed .global() table exactly as it does on D1: same tokenizer, same matching rules, same relevance order (see Full-text search). Neither Postgres nor MySQL ships SQLite's FTS5, so the store maintains a portable inverted companion instead: one indexed (token, document, occurrences) row per distinct token, updated with each row write and read back in a single indexed query. On Postgres its token index declares text_pattern_ops so the prefix match of a query's final term stays indexed under any collation; on MySQL both columns take the InnoDB key prefix.

Rows that predate the index are backfilled a bounded page per request, with progress recorded in __lunora_search_state, so a large table becomes searchable progressively rather than stalling the first request after a deploy. Pass staged: true to keep that work out of the request path entirely and drive it yourself with backfillSqlSearchIndexes.

Ranking aggregates over every matching token row, so the 1024-document limit bounds what you get back, not what the database reads. On Postgres you can trade that away with .searchIndex({ strategy: "native" }), which stores a tsvector per document and lets a GIN index answer the match: same documents, ordered by the engine rather than by Lunora's scorer.

The convenience constructors cover the common path. For custom wiring, the lower-level pieces are also exported:

  • buildPgExec(client) / buildMysqlExec(load) turn a driver into the store's SqlExec.
  • postgresDialect / mysqlDialect are the engine dialects.
  • createHyperdriveGlobalCtxDb({ engine, exec, ...storeOptions }) is the general factory the two convenience constructors call.

Vector search on your own Postgres

.vectorize() normally writes through to a Cloudflare Vectorize binding. If you have already moved the relational tier onto your own Postgres, that split is awkward: full-text search runs natively (see strategy: "native" above) while vectors alone still reach for a Cloudflare service.

createPgVectorIndex closes it. The shard takes vectors: (env) => Record<string, VectorizeIndexLike>, and everything above that — the write-through sync hook, ctx.vectors, codegen — is written against the interface rather than the provider, so a pgvector index drops in where the binding went:

src/server/index.ts
import { createHyperdrive, fromPostgresJs } from "@lunora/hyperdrive";
import { createPgVectorIndex } from "@lunora/hyperdrive/global";

// Built once, not per request: `vectors` is called on every dispatch, and a fresh
// index object means a fresh provisioning memo — five extra round trips each time.
let index: ReturnType<typeof createPgVectorIndex> | undefined;

createShardDO({
    vectors: (env) => {
        index ??= createPgVectorIndex({
            client: fromPostgresJs(async () => {
                const { default: postgres } = await import("postgres");

                return postgres(createHyperdrive((env as Env).HYPERDRIVE).connectionString, { fetch_types: false, max: 5 });
            }),
            dimensions: 768,
            metric: "cosine",
            name: "posts_search",
        });

        return { posts_search: index };
    },
});

Nothing else changes: .vectorize() in the schema, ctx.vectors.query(...) in a function, and the automatic write-through sync all behave the same.

The table (__vec_<name>), the vector extension, an HNSW index in the metric's operator class, and btree/GIN indexes on the two filterable columns are provisioned lazily on first use — every statement is IF NOT EXISTS, so it is safe against a live table. The connecting role needs to be able to run CREATE EXTENSION vector; on a managed Postgres where it cannot, install the extension once out-of-band and the statement becomes a no-op.

Four deliberate differences from Vectorize. Metadata filters are JSONB containment — equality for scalars, but for containers it is containment, not equality: { tags: ["x"] } also matches a row whose tags are ["x", "y"], and { author: {} } matches every row carrying an author key. A filter value JSON cannot represent (undefined, a function, a symbol) is rejected, because dropping it would widen the filter to match every row. A comparison operator like { views: { $gt: 10 } } and Vectorize's dot-addressed "author.role" syntax throw for the same reason. returnMetadata: "indexed" behaves as "all", since Postgres indexes the whole JSONB document — note this defeats a deliberate default: ctx.vectors asks for "indexed" precisely so a query that opted into nothing does not return every stored field. And dimensions is fixed at first provisioning: CREATE TABLE IF NOT EXISTS no-ops on a later change, so writes fail against the original width. Finally mutationId is a local counter: Vectorize's is a handle for an async mutation queue, while a Postgres write is already durable when its promise resolves.

Two ceilings worth knowing. dimensions is capped at 2000, pgvector's HNSW limit — rejected at construction rather than failing opaquely at CREATE INDEX (Vectorize itself tops out at 1536, so most ports are unaffected). And name is capped at 34 characters, because the longest index name derived from it (__vec_ + name + __ann_ + the operator class) has to stay under Postgres' 63-byte identifier truncation — past that, two indexes would collapse onto one name and the second CREATE INDEX IF NOT EXISTS would silently no-op.

Pass the same metric the schema's .vectorize(...) declares. Nothing cross-checks the two: mismatch them and the index is built for one operator and queried with another, so the ordering is wrong for the declared metric and every score is on a different scale — with no error at any layer.

The indexes on namespace and metadata are not an optimisation. pgvector applies WHERE after scanning the ANN index, so without a usable index on the filtered column the planner's only path is the approximate one, and a filter matching a small slice of the table returns a short page with no error. Indexing both lets the planner answer the filter exactly instead. (hnsw.iterative_scan is pgvector's own answer to this, but it is a session setting, and a pooled client can run the SET and the query on different sessions.)

metric picks the distance operator and the matching index operator class — cosine (default), euclidean, or dot-product. score is reported the way Vectorize reports it: cosine similarity (1 = identical), raw L2 distance (0 = identical), or the NEGATIVE inner dot product — Cloudflare's own convention, where a score of -1000 is more similar than -500.

Non-goals

  • No CDC / logical replication. Lunora does not ingest your Postgres write-ahead log; the projection pattern above is the supported path to reactivity.
  • No ctx.sql in query/mutation. Enforced by the hyperdrive_outside_action advisor lint.
  • No bundled driver / ORM. You own driver choice and lifecycle.

Public API

ExportPurpose
createHyperdrive(binding)Lift connectionString + discrete parts off the binding
fromPostgresJs(load)Wrap a postgres.js client, built by load on first query, as a SqlClient
fromNodePg(load)Wrap a pg Client/Pool, built by load on first query, as a SqlClient
fromMysql2(load)Wrap a mysql2/promise connection/pool, built by load, as a SqlClient
pullSourceRows(sql, { query, params, idColumn?, map? })Run a tenant query + project rows to Lunora docs (read side of the ingest bridge)
projectSourceRow(row, { idColumn?, map? })Project one external row → a document with _id
SqlClient, HyperdriveLike, HyperdriveConnection, PostgresJsLike, NodePgLike, Mysql2Like, ProjectOptions, PullSourceOptionsType-only

From @lunora/hyperdrive/global (reactive .global() backend):

ExportPurpose
createPostgresGlobalCtxDb(client, options)Build a reactive Postgres .global() writer (globalDb)
createMysqlGlobalCtxDb(load, options)Build a reactive MySQL .global() writer (needs FOUND_ROWS)
createHyperdriveGlobalCtxDb({ engine, exec, … })General factory the two convenience constructors call
buildPgExec(client) / buildMysqlExec(load)Wrap a driver as the store's SqlExec
postgresDialect / mysqlDialectThe engine dialects
createPgVectorIndex({ client, name, dimensions })A pgvector-backed index for ctx.vectors (no Vectorize binding); name is the SQL table identifier, not the vectors map key
exportHyperdriveGlobalRows / importHyperdriveGlobalRowsThe admin export/import of .global() rows (lunora export/import, migrate d1-to-hyperdrive and back); codegen wires them
createHyperdriveGlobalIntrospector(engine, exec, schema)The Studio's global data browser over the Hyperdrive database; codegen wires it
readHyperdriveGlobalCdc / applyHyperdriveGlobalCdcThe global changelog read (admin CDC sync) and replay (point-in-time apply); codegen wires them

See also