Server mode

No — this does not replace the mode you already use. Embedded/local mode — one process, no daemon, straight to SQLite — stays the default and stays why you'd pick FlexiQ over a broker-backed queue. Server mode is additive: a separate flexiq-server process that holds the database credential and exposes it over the network, for the cases where nothing in-process can reach that credential at all.

Writing that caller is its own page, one per door: Clients without an SDK for the producer side, Custom executors for the executor side.

Nothing forces an all-or-nothing switch. A deployment can put only its producer on the network door and leave its workers opening storage directly, same as today — the "Try it" section below points at an example of exactly that hybrid.

What it costs

The prefork pool, and the worker pool underneath it, assume an OS process on the same machine as the jobs it runs — that's where the CPU-parallelism story lives. A remote executor, whether it dials in over the attach protocol today or over flexiq.executor.v1 once a client speaks it, gets none of that: it is a Celery-style remote consumer, one network hop from the scheduler, with the capacity of whatever container it happens to run in. Server mode buys you a network door; it does not buy you more cores.

The two doors

flexiq-server puts two independent things on the network, not one door with two names:

A produce token cannot open an executor stream, and an execute token cannot enqueue — they are not a hierarchy, and a credential scoped to one package never opens the other. See the gRPC door for the producer RPCs and the token lifecycle, and the executor door for the Attach stream.

raw versus structured

The producer door takes a job's payload two ways. raw is the CBOR envelope an SDK already builds, byte for byte — nothing lost, but it needs a CBOR library. structured takes ordinary JSON values and lets the server build that same envelope, so a shell script can enqueue with no codec at all. What it cannot carry, it refuses rather than rounds: integers past 2^53 - 1, non-finite numbers, and anything JSON has no syntax for. See the full treatment for the one thing it silently normalises instead — object key order — and what that costs an auto: idempotency key.

Try it

examples/polyglot has a gRPC-speaking variant beside the original shared-database one: a producer that enqueues over the gRPC door with grpcurl, no SDK installed at all, and a Node worker that attaches to the scheduler rather than opening the file. That worker also shows what an executor gives up — it fans out to a second stage, and with no storage to write the follow-up job into, it sends it back through the same producer door. The Java worker still opens the database file directly.

Trade-offs