What is a Semaphore in Java and what problem does it solve?

## The problem Imagine a shared resource that can only safely handle a limited number of users at the same time — say a connection pool with 5 connections, or a downstream service that allows at most 10 concurrent calls. If 100 threads rush in, you need a gatekeeper that lets only N through and makes the rest wait. A **mutex** (mutual-exclusion lock) only allows *one* at a time. A **Semaphore** generalises this to *N* at a time. ## Permits A `Semaphore` is created with a number of **permits**: `new Semaphore(5)` starts with 5. Think of permits as physical tickets in a bowl: - **`acquire()`** removes one ticket from the bowl. If the bowl is empty, the calling thread **blocks** (waits, doing nothing, not spinning) until another thread puts a ticket back. - **`release()`** puts one ticket back into the bowl, and if any thread is waiting in `acquire()`, one of them is woken up and proceeds. The count of available permits goes up and down as threads acquire and release. When it hits zero, the next `acquire()` waits. ## The standard pattern ```java Semaphore sem = new Semaphore(5); sem.acquire(); // take a permit (may block) try { useTheResource(); // at most 5 threads are here at once } finally { sem.release(); // ALWAYS give it back } ``` The `release()` belongs in a `finally` block so the permit is returned even if the work throws an exception. Forgetting this **leaks** a permit — the bowl permanently shrinks, and eventually everyone blocks forever. ## No owner — the key difference from a lock A `ReentrantLock` or `synchronized` block has an **owner**: only the thread that locked it may unlock it, and it is **reentrant** (the owner can re-lock without deadlocking). A `Semaphore` has **neither**: - Any thread may call `release()`, even one that never called `acquire()`. This is intentional — it lets you use a semaphore as a signal between producer and consumer threads. - It is **not reentrant**: if the same thread calls `acquire()` twice on a `Semaphore(1)`, it deadlocks against itself. - You can `release()` more than you acquired, which *increases* the permit count above the starting value — a legitimate way to add capacity to a pool at runtime. Because of this, a `Semaphore` is *not* a drop-in lock for protecting a critical section where you need reentrancy or owner checks. A `Semaphore(1)` is a **binary semaphore** that behaves like a non-reentrant mutex, useful mainly when you need cross-thread signalling. ## Counting vs binary - **Counting semaphore**: permits > 1 — bounds concurrency to N (pool limiting, rate gating). - **Binary semaphore**: 1 permit — mutual exclusion or a single on/off signal. ## Non-blocking and timed acquisition `tryAcquire()` attempts to take a permit and returns `true`/`false` *immediately* without blocking — useful to shed load ("if no capacity, fail fast"). `tryAcquire(timeout, unit)` waits up to a bound. You can also acquire/release **multiple permits at once**: `acquire(3)` / `release(3)`. ## Fairness The constructor takes an optional boolean: `new Semaphore(5, true)` makes it **fair** — waiting threads are granted permits in FIFO (first-come-first-served) order. The default (`false`) is **non-fair**: a newly arriving thread may "barge" and grab a just-released permit ahead of threads that have been waiting longer. Non-fair has higher throughput (fewer context switches); fair prevents **starvation** (a thread waiting indefinitely) at some throughput cost. ## Under the hood `Semaphore` is built on `AbstractQueuedSynchronizer` (AQS), the same framework behind `ReentrantLock` and `CountDownLatch`. The permit count is AQS's shared integer state, and waiting threads sit in AQS's FIFO wait queue. ## When to reach for it Resource pools, throttling concurrent requests to a service, implementing a bounded buffer's capacity, or simple cross-thread signalling. When you just need exclusive access to a critical section, prefer `synchronized` or `ReentrantLock` (reentrant, owner-checked); use a semaphore when the *count* of concurrent holders matters.