+</pre><h2><a name="Design pattern: Implementing locking with SETNX">Design pattern: Implementing locking with SETNX</a></h2><blockquote>SETNX can also be seen as a locking primitive. For instance to acquirethe lock of the key <b>foo</b>, the client could try the following:</blockquote>
+<pre class="codeblock python python" name="code">
+SETNX lock.foo <current UNIX time + lock timeout + 1>
+</pre><blockquote>If SETNX returns 1 the client acquired the lock, setting the <b>lock.foo</b>key to the UNIX time at witch the lock should no longer be considered valid.The client will later use <b>DEL lock.foo</b> in order to release the lock.</blockquote>
+<blockquote>If SETNX returns 0 the key is already locked by some other client. We caneither return to the caller if it's a non blocking lock, or enter aloop retrying to hold the lock until we succeed or some kind of timeoutexpires.</blockquote>
+<h3><a name="Handling deadlocks">Handling deadlocks</a></h3><blockquote>In the above locking algorithm there is a problem: what happens if a clientfails, crashes, or is otherwise not able to release the lock?It's possible to detect this condition because the lock key contains aUNIX timestamp. If such a timestamp is <= the current Unix time the lockis no longer valid.</blockquote>
+<blockquote>When this happens we can't just call DEL against the key to remove the lockand then try to issue a SETNX, as there is a race condition here, whenmultiple clients detected an expired lock and are trying to release it.</blockquote>
+<ul><li> C1 and C2 read lock.foo to check the timestamp, because SETNX returned 0 to both C1 and C2, as the lock is still hold by C3 that crashed after holding the lock.</li><li> C1 sends DEL lock.foo</li><li> C1 sends SETNX => success!</li><li> C2 sends DEL lock.foo</li><li> C2 sends SETNX => success!</li><li> ERROR: both C1 and C2 acquired the lock because of the race condition.</li></ul>
+<blockquote>Fortunately it's possible to avoid this issue using the following algorithm.Let's see how C4, our sane client, uses the good algorithm:</blockquote>
+<ul><li> C4 sends SETNX lock.foo in order to acquire the lock</li><li> The crashed C3 client still holds it, so Redis will reply with 0 to C4.</li><li> C4 GET lock.foo to check if the lock expired. If not it will sleep one second (for instance) and retry from the start.</li><li> If instead the lock is expired because the UNIX time at lock.foo is older than the current UNIX time, C4 tries to perform GETSET lock.foo <current unix timestamp + lock timeout + 1></li><li> Thanks to the <a href="GetsetCommand.html">GETSET</a> command semantic C4 can check if the old value stored at key is still an expired timestamp. If so we acquired the lock!</li><li> Otherwise if another client, for instance C5, was faster than C4 and acquired the lock with the GETSET operation, C4 GETSET operation will return a non expired timestamp. C4 will simply restart from the first step. Note that even if C4 set the key a bit a few seconds in the future this is not a problem.</li></ul>
+IMPORTANT NOTE: In order to make this locking algorithm more robust, a client holding a lock should always check the timeout didn't expired before to unlock the key with DEL because client failures can be complex, not just crashing but also blocking a lot of time against some operation and trying to issue DEL after a lot of time (when the LOCK is already hold by some other client).
projects / redis.git / blobdiff