Is a binary semaphore always the same as a mutex?

They can serve similar exclusion purposes, but ownership and release rules may differ by implementation.

Does a semaphore of five enforce five calls per second?

No. It limits simultaneous permit holders; a rate limit needs time-based logic.

Can it prevent every race condition?

No. Shared data still needs appropriate synchronization, and code outside the semaphore protocol can violate the intended limit.

AI Glossary / Data engineering and architecture / Intermediate

Semaphore (programming)

Definition

A semaphore is a synchronization mechanism that controls access to a limited number of permits or signals the availability of something between concurrent activities. A counting semaphore maintains a count: acquiring consumes a permit when one is available, and releasing adds one. If no permit is available, an acquire operation may wait or fail according to the interface used. Semaphores help coordinate concurrency but do not automatically make an entire program correct.

What it means

Suppose a service can safely run three expensive jobs at once. A semaphore initialized with three permits can limit how many cooperating workers enter that part of the program simultaneously. Other workers wait until a permit becomes available. This is a concurrency limit, not necessarily a limit on requests per second. Fast jobs can pass through the three slots many times in a short interval.

How it works

Create the semaphore with an appropriate initial count. Before using the controlled resource, acquire a permit; after the operation ends, release it through a reliable cleanup path. An implementation typically makes the counter operations atomic. Waiting, timeout, fairness, and ownership behavior depend on the API. A bounded semaphore can help detect releases that exceed its initial capacity in some libraries. All participating code must follow the protocol, and failures or cancellation must not leave permits permanently consumed.

Three practical examples

Connection access: an application limits simultaneous use of a small pool of external connections, ensuring that workers wait rather than exceed the planned capacity.
AI inference: a service restricts concurrent memory-intensive model calls to avoid launching more work than its resource budget can support.
Producer-consumer coordination: one activity signals that an item is available and another waits for that signal, alongside a separately protected data structure that stores the item.

Think of it this way

Think of a room with three admission tokens. A visitor takes one before entering and returns it on leaving; without a token, the next visitor waits. The analogy is limited because software must handle exceptions, cancellation, and scheduling. A forgotten return can block future work, while an extra return can falsely suggest more capacity than exists. The token system also does not control what visitors do to shared objects inside.

Worked example

A semaphore starts at 3. Workers A, B, and C each acquire once, reducing available permits to 0. Worker D attempts to acquire and waits. A finishes and releases, allowing D to proceed while B and C continue. At most three cooperating workers occupy the controlled section. If B raises an exception and its cleanup fails to release the permit, effective capacity may fall to two after the other workers finish. If C releases twice, an unbounded counting semaphore may permit more work than intended. The example explains why acquire and release must be paired correctly and why a concurrency test should include failure paths.

Illustrative case study

Fictional example: Maple Images processes uploaded photographs using a memory-heavy operation. Engineers place a semaphore around that operation and test cancellation, errors, and normal completion. They discover that one early-return path skips cleanup and fix it before relying on the limit. They also protect shared bookkeeping separately because the semaphore allows several workers at once. This invented case illustrates coordination design; it does not claim a particular capacity or performance result for an actual image-processing service.

Common misconceptions and mistakes

Do not treat a counting semaphore as automatic mutual exclusion when it allows more than one entrant. Do not assume waiting workers are served in first-in-first-out order unless the implementation guarantees it. Avoid confusing concurrency limits with rate limits. Also distinguish a semaphore from a mutex: mutexes commonly express exclusive ownership, while semaphore operations can support different signaling patterns. The exact rules vary by API, so code should not transfer assumptions blindly between libraries or operating systems.

Business use

Semaphores are useful when a service needs explicit coordination around scarce resources. Choose a limit based on measured resource behavior and downstream constraints, then monitor wait time and active work. Include timeout and cancellation handling so waiting does not become an invisible backlog. For AI workloads, a semaphore can help bound simultaneous model calls, but queue size, memory usage, retries, and per-request cost still need their own controls. Test that every exit path restores the intended permit count.

Check your understanding

A worker acquires a permit and then fails before releasing it. Why can unrelated later work stall?

Answer: The missing release leaves a permit consumed. Repeated failures can exhaust available permits even when no useful work is running, so cleanup must cover exceptional paths.

Related terms

Mutex
Synchronization
lock
concurrency
bounded semaphore
producer-consumer problem
rate limiting

Learn next

trace a three-permit semaphore with five workers on paper. Include normal completion, timeout before acquisition, and failure after acquisition. Mark exactly which paths must release a permit and which must not release one they never acquired.

Sources

Primary source consulted: Python threading documentation, https://docs.python.org/3.13/library/threading.html — Semaphore and BoundedSemaphore search excerpts consulted. The entry explains a general coordination concept; Python-specific details are examples rather than universal API guarantees. Worked sequences and fictional case are original. Full editorial review remains pending.

Previous: Semantometrics Next: Semi-amortized variational inference