Simpy Python

SimPy Python

Build reproducible discrete-event models whose clock, state transitions, resource ownership, stopping rule, randomness, and observations represent the real system rather than an accidental execution order.

Boundary and core objects

A process function is a generator recipe; env.process(generator) registers it and returns a Process event. yield event suspends simulated activity until that event is processed. Calling a generator normally does not run the model. Read the event and process model before mixing callbacks, interrupts, conditions, or resource requests.

Ordered workflow

1. Define the question, time unit, warm-up, horizon/stopping event, and measured population before modeling implementation details.
2. Separate domain state from the environment and random-number generator.
3. Express every delay, wait, acquisition, release, failure, and cancellation as an event transition.
4. Select Resource, Container, or Store from what is constrained—not by superficial queue vocabulary.
5. Record observations at defined transitions and with clear denominators.
6. Run independent replications with controlled seeds; report uncertainty, not one trace, when making stochastic claims.
7. Test tiny deterministic timelines, contention, boundary times, interruption, empty/full queues, and the stopping rule.

Decision rules

• Use Resource when entities compete for interchangeable usage slots. Use a with resource.request() as request: yield request block so ownership is released on normal exit or interruption.
• Use PriorityResource when waiting order follows an explicit priority. State whether lower numeric values mean higher priority and define tie behavior.
• Use PreemptiveResource only when an admitted user can be interrupted. Handle simpy.Interrupt, remaining work, cleanup, and measurement bias.
• Use Container for an aggregate quantity without item identity. Use Store when objects and selection/filtering matter.
• Use env.timeout(duration) for simulated time. Reject negative durations and define whether zero delay is a legitimate same-time event.
• Use env.any_of/event1 | event2 for races and all_of/& for barriers. Inspect which events completed; a losing event may still need cancellation or cleanup.
• Use env.run(until=t) knowing events scheduled exactly at t are not a vague “through time t” promise; write a boundary test for the chosen stopping form.
• Use RealtimeEnvironment only when coupling to wall-clock behavior is an explicit requirement. It is not a way to make simulation more accurate.

Read resources, time, and experiments for ownership, queues, monitoring, randomness, and replication rules.

Canonical anchor

from collections.abc import Generator
from dataclasses import dataclass

import simpy


@dataclass(frozen=True)
class Visit:
    customer: int
    queued_at: float
    started_at: float
    finished_at: float


def customer(
    env: simpy.Environment,
    server: simpy.Resource,
    customer_id: int,
    service_time: float,
    visits: list[Visit],
) -> Generator[simpy.Event, object, None]:
    queued_at = env.now
    with server.request() as request:
        yield request
        started_at = env.now
        yield env.timeout(service_time)
        visits.append(Visit(customer_id, queued_at, started_at, env.now))

The record captures state-transition times rather than sampling a mutable queue. Arrival creation and random service draws belong outside this process so a test can inject a deterministic schedule.

High-risk rules

• Do not mutate state “after a delay” without yielding that delay.
• Do not hold a resource while waiting for unrelated work unless the real entity occupies it during that time.
• Do not manually release a context-managed request twice.
• Never share one mutable RNG implicitly across replications. Derive and record per-replication seeds; common random numbers require an intentional design.
• Avoid global random calls inside processes. Inject distributions or an RNG.
• A queue-length time average requires time weighting between changes; averaging observations at arrivals answers a different question.
• Distinguish censored entities still in the system at the horizon from completed entities. Do not silently discard them from denominators.
• Same-time events are ordered by scheduler rules and creation sequence. If the answer must not depend on that order, encode priorities or redesign the event.
• An interrupted process must release resources and decide whether work resumes, restarts, or is lost. Catching and ignoring the interrupt is not a model.

Version grounding and completion

The authoring baseline is official SimPy 4.1.2 stable documentation; the local foundry has not yet executed the package. Check the installed version with simpy.__version__ and inspect signatures before relying on exception fields, resource variants, or environment behavior. Read verification.

Completion requires an explicit clock/stopping contract, correct resource ownership, injected randomness, traceable observations, deterministic micro-tests, stochastic replication checks where relevant, and evidence that the result does not depend accidentally on global state or one seed.

References

• Event and process model
• Resources, time, and experiments
• Verification and grounding