monty-pool

A pool of Monty worker processes for running untrusted Python code with crash isolation.

Monty executes untrusted Python, and a Monty process can never be made fully crash-proof against memory errors (stack overflow aborts, allocator aborts). This crate isolates those crashes by running the interpreter only in worker subprocesses, reached over Monty's wire protocol: a crashed worker kills only itself, the pool detects the death and replaces the worker, and the parent process is never at risk.

This is the recommended way to run Monty from Rust. It is also the engine underneath the pydantic-monty Python package and the @pydantic/monty JavaScript package.

Model

A Pool keeps an elastic set of workers (min_processes prewarmed, up to max_processes). Pool::checkout dedicates one worker to one REPL session: the caller feeds snippets of code and answers suspension events (TurnEvent — external function calls, OS calls, name lookups, async futures) until the snippet completes, then Checkout::finish returns the worker to the pool for reuse. A Checkout dropped without finish kills its worker instead — mid-execution state cannot be trusted back into the pool.

The pool is async end-to-end and runs on tokio: frame reads are cancel-safe (partial-frame state lives in the worker), and turn deadlines are ordinary timers rather than a watchdog thread. Turn futures are not resumable after being dropped mid-flight — the checkout notices, discards the worker, and fails the next call cleanly.

Usage

Workers are monty CLI binaries spawned as subprocesses — build one with cargo build -p monty-runtime in the Monty repository, or install it from PyPI as pydantic-monty-runtime.

use std::time::Duration;

use monty_pool::{Pool, PoolConfig, PoolError, ReplConfig, TurnEvent, on_print_sync};

#[tokio::main]
async fn main() -> Result<(), PoolError> {
    let mut config = PoolConfig::subprocess("path/to/monty");
    // no timeouts by default; set one before running untrusted code
    config.request_timeout = Some(Duration::from_secs(30));
    let pool = Pool::new(config).await?;

    let mut session = pool.checkout(&ReplConfig::default()).await?;
    let mut on_print = on_print_sync(|_stream, text| print!("{text}"));

    // session state persists between feeds on the same checkout
    session.feed("x = 21", vec![], vec![], false, &mut on_print).await?;
    let event = session.feed("x * 2", vec![], vec![], false, &mut on_print).await?;
    match event {
        TurnEvent::Complete(value) => println!("result: {value:?}"), // Int(42)
        // other events are suspensions (external function calls, OS calls,
        // name lookups, futures) answered with `resume` / `resume_name_lookup`
        // / `resume_futures` to continue the turn
        other => println!("suspended: {other:?}"),
    }

    // return the worker to the pool for reuse by the next checkout
    session.finish().await?;
    Ok(())
}

ReplConfig also enables per-session sandbox ResourceLimits and type checking of every fed snippet; Checkout::feed accepts inputs (host values exposed as sandbox globals) and per-feed filesystem mounts (MountSpec). Sessions can be snapshotted with Checkout::dump and restored later — including on a different worker or machine — with Checkout::restore.

Protections over in-process execution

• Crash isolation — a segfault, stack-overflow abort, or allocator abort in the sandbox kills only the worker. The pool observes the death as PoolError::Crashed, discards the worker, and spawns a replacement; the parent process and every other session stay healthy.
• Hard timeouts — a parent-side deadline kills any worker whose turn exceeds request_timeout (PoolError::Timeout), backstopping the sandbox's own resource limits and catching hangs those limits cannot see. Synchronous host telemetry processors delay enforcement while they run because the timer cannot be polled. When a session has a max_duration budget, the deadline also enforces it (plus duration_limit_grace) from outside the child. A max_suspensions budget is enforced by the pool alone: it counts the suspensions it services and ends the feed past the budget with an uncatchable RuntimeError in the sandbox. PoolConfig::subprocess sets neither request_timeout nor checkout_timeout by default; set request_timeout yourself for untrusted code.
• Untrusted children — the parent treats every frame from a (possibly compromised) worker as untrusted: wire decoding validates everything and never panics, and a worker that violates the protocol is discarded.
• Worker recycling — max_checkouts_per_worker recycles long-lived children to bound the impact of any slow leak.
• Memory limits — a session's max_memory also caps the worker's live allocations, enforced in the worker's own global allocator (monty-alloc) plus 4 MB of headroom (32 MB with type checking), rather than letting a worker grow the host until the OOM killer intervenes. Exceeding it, or a refused allocation, exits the worker with a dedicated code so it is reported as PoolError::Runtime/MemoryError instead of an unclassifiable abort — the one Runtime error whose worker does not survive.

Runtime errors inside the sandbox (PoolError::Runtime) are not crashes: the worker and its session remain alive and usable — the one exception being the MemoryError above, raised for a worker that has already exited.

Observability

The optional telemetry feature records semantic execution for language bindings and other hosts. The former telemetry-adapter feature and telemetry_adapter module remain compatibility aliases. monty-pool never selects or shuts down a network exporter or reads exporter credentials; the host SDK owns those choices and its final flush/shutdown.

Recording happens in the host process, which builds every request and decodes every event anyway, so both transports are covered and the workers stay uninstrumented. Each instrumented checkout becomes one session span; each feed is a nested span held across suspension round-trips, with a child span per suspension whose duration is the host round-trip. Fed code, inputs, call arguments and results, exceptions and print output are recorded in full — values encoded the way the Python logfire SDK encodes attributes, capped at 64KB per value — while Load/Dump snapshot blobs are recorded by size only. Supplying an SDK is therefore an explicit opt-in to recording potentially sensitive values.

The adapter configures a process-global Rust pipeline and returns a handle that creates each checkout's serialized parent context. Span and log records are emitted through TelemetryAdapter; metrics are aggregated by that pipeline and exported through the adapter as standard OTLP protobuf batches. Python, Node, and third-party bindings retain ownership of their native SDK and exporter. Without the feature, workers contain no telemetry recorder or telemetry hot path.

Metrics

The same handle yields a Metrics for PoolConfig::metrics — as does Metrics::for_logfire for a Rust host. Both record into real Logfire instruments, with exponential histogram buckets; the adapter pipeline differs only in exporting its resulting aggregate over the language boundary. Either turns on the aggregate side: pool health (monty.pool.workers.live, monty.pool.workers.idle, monty.pool.workers.suspended, monty.pool.checkout.wait, monty.pool.worker.terminated, monty.pool.session.duration) and per-turn cost (monty.run.duration, monty.run.execution_time, monty.turn.duration, monty.run.suspensions, monty.ext.call.duration, monty.snapshot.bytes, monty.print.bytes, monty.wire.frame.bytes). monty.pool.session.duration uses ok for a clean finish, error when the worker is lost, and abandoned when a live checkout is dropped.

Two differences from the spans above. Metrics cover every checkout, not only the ones a host gave a parent context — an aggregate over traced sessions alone would be misleading — and they record no sandbox-supplied values at all: every attribute is a closed set, so a called function's name (under any outcome — a host lookup that is a callable resolves anything), an exception class and any path are all left out rather than becoming a time series each. The one name recorded is an os call's, which comes from the protocol's own fixed set. The subtraction worth knowing: monty.run.duration minus monty.run.execution_time is host and transport overhead, primarily time spent answering suspensions.

Metric attributes deliberately never identify a pool. The worker up/down counters therefore total over all pools recording into the same host meter, and a dropped pool subtracts its remaining contribution.

A foreign SDK receives aggregated ExportMetricsServiceRequest protobufs through TelemetryAdapter::export_metrics, which defaults to dropping them so a span-only adapter continues to work. Its flush path should call TelemetryAdapterHandle::force_flush before flushing the host exporter.

Transports

• Subprocess (PoolConfig::subprocess) — spawn local monty subprocess children over framed stdio. These are the poolable workers: prewarmed, reused across checkouts, and replaced on crash.
• WebSocket (PoolConfig::websocket) — dial a remote child (or a relay pairing the two ends) over ws:///wss://. These workers are single-use: dialed fresh per checkout, never prewarmed or returned to the pool. Isolation is the remote host's responsibility — a remote crash is observed as the connection dropping. Pool::checkout_with takes CheckoutOptions::connect_headers, extra headers for that checkout's upgrade request — e.g. a token for a relay in front of the worker. The request carries User-Agent: monty-pool/<version>, and with the telemetry feature the traceparent (and tracestate) of CheckoutOptions::telemetry, so server-side spans join the caller's trace; a connect_headers entry of the same name replaces either.

Monty crates

• monty — the core interpreter: Python parser, bytecode VM, and sandbox.
• monty-types — the shared boundary data types (values, exceptions, OS calls, resource limits) hosts use without linking the interpreter.
• monty-fs — host-side filesystem mounts: maps virtual sandbox paths to real host directories.
• monty-runtime — the monty binary: REPL, file runner, and subprocess worker mode.
• monty-pool — an elastic pool of crash-isolated monty worker subprocesses. this crate
• monty-proto — the protobuf wire protocol spoken between pool parents and workers.
• monty-type-checking — type checking of sandboxed code, powered by ty.
• monty-typeshed — the trimmed typeshed stubs describing the stdlib subset Monty implements.
• monty-macros — the proc macros behind monty's argument parsing.