Blog / Product · · 3 min
Why Akter
Stateful backends break in the seams between databases, queues, and sandboxes. Akter gives every resource an Effect actor with SQLite, POSIX files, workflows, and microVMs.
Stateful backends usually split one domain model across five infrastructure systems. Teams run Postgres for rows, Redis for locks, S3 for files, Temporal for workflows, and container runners for sandboxes. Each service fails on a separate clock, so production bugs hide in the seams between them. Replies leave before writes commit, retries charge cards twice, and rolling deploys drop active streams.
Every resource gets one sequential actor
Akter models your backend around the resources your product actually owns. An order, a ledger, a collaborative room, a workflow, or a sandbox workspace each becomes an actor created with Akter.make.
Every actor obeys strict Actor-model bounds. Only that actor reads or writes its private SQLite database and POSIX fs directory. One sequential writer processes state turns one at a time, so handlers never race or need distributed locks. Hosts send replies, push WebSocket updates, and start jobs only after the SQLite savepoint commits.
Slow or fallible external work runs in step-journaled jobs outside the SQLite savepoint. Each finished activity records its outcome in the actor database, so resumed workflows never repeat completed work. Actors also own HTTP and SSE routes, zero-disk WebSocket presence channels, timers, cron schedules, and Firecracker Linux microVMs.
Built natively on Effect
Akter builds directly on Effect through @akterjs/effect. Contracts declare states, inputs, outputs, and tagged errors with Schema, while handlers compose SqlClient, FileSystem, Stream, and Layer. Because the contract types the entire boundary, clients and Story tests catch schema and error mismatches at compile time.
One akterd engine on your laptop and in Akter Cloud
akterd is the Rust actor engine behind every Akter app. In Akter Cloud, akterd runs clustered across three AWS availability zones on c8id NVMe hosts. Each turn commits to a 2-of-3 NVMe quorum WAL, and cold segments tier to R2. Computers run as Firecracker microVMs on the same hosts, and Cloudflare Anycast routes mTLS traffic straight to them.
@akterjs/cli ships akterd as a stripped precompiled binary for aarch64 and x86_64. akter dev and offline tests run that binary with --local. You install one package and run one binary with no Docker. The published @akterjs/effect package holds declarative contract schemas, client stubs, and the TurnWorker loop.
Your data stays in your bucket
akter backup continuously streams standard .sqlite snapshots and POSIX .tar.zst archives into your own S3 or R2 bucket. You choose the schedule. Stock sqlite3 and tar open every file without Akter installed. akter export writes the same files on demand.
Architectural boundaries and tradeoffs
- Cross-actor writes use explicit coordination.
Akter.transactionruns each participant actor's own command atomically in sorted lock order rather than exposing raw cross-actor SQL. - Cross-actor reads use projections. Queries read one actor database at a time, while
Akter.viewbuilds read-only SQLite projections from shard commit logs. - External side effects live in jobs. Commands cannot make outbound network calls inside a SQLite savepoint, so external API calls run inside recorded job activities.
- Computers never mount the database. A Linux microVM mounts the actor
fsfolder for workspace files, whileactor.dbstays private on the host.
Run locally, deploy, and back up
- To run an app locally on
akterdwith zero cloud setup, runnpx @akterjs/cli dev. - To deploy to managed multi-AZ infrastructure, run
npx @akterjs/cli deploy --organization acme --app orders. - To stream backups into your own bucket, follow the backup guide.