</pre>
Imagine a Redis version that does not implement the "Delete keys with an expire set on write operation" semantic.
Running the above example with the 10 seconds pause will lead to 'a' being set to the value of 1, as it no longer exists when INCR is called 10 seconds later.<br/><br/>Instead if we drop the 10 seconds pause, the result is that 'a' is set to 101.<br/><br/>And in the practice timing changes! For instance the client may wait 10 seconds before INCR, but the sequence written in the Append Only File (and later replayed-back as fast as possible when Redis is restarted) will not have the pause. Even if we add a timestamp in the AOF, when the time difference is smaller than our timer resolution, we have a race condition.<br/><br/>The same happens with master-slave replication. Again, consider the example above: the client will use the same sequence of commands without the 10 seconds pause, but the replication link will slow down for a few seconds due to a network problem. Result? The master will contain 'a' set to 101, the slave 'a' set to 1.<br/><br/>The only way to avoid this but at the same time have reliable non time dependent timeouts on keys is to destroy volatile keys when a write operation is attempted against it.<br/><br/>After all Redis is one of the rare fully persistent databases that will give you EXPIRE. This comes to a cost :)<h2><a name="FAQ: How this limitations were solved in Redis versions > 2.1.3?">FAQ: How this limitations were solved in Redis versions > 2.1.3?</a></h2>Since Redis 2.1.3 there are no longer restrictions in the use you can do of write commands against volatile keys, still the replication and AOF file are guaranteed to be fully consistent.<br/><br/>In order to obtain a correct behavior without sacrificing consistency now when a key expires, a DEL operation is synthesized in both the AOF file and against all the attached slaves. This way the expiration process is centralized in the master instance, and there is no longer a chance of consistency errors.<br/><br/>However while the slaves while connected to a master will not expire keys independently, they'll still take the full state of the expires existing in the dataset, so when a slave is elected to a master it will be able to expire the keys independently, fully acting as a master.
</pre>
Imagine a Redis version that does not implement the "Delete keys with an expire set on write operation" semantic.
Running the above example with the 10 seconds pause will lead to 'a' being set to the value of 1, as it no longer exists when INCR is called 10 seconds later.<br/><br/>Instead if we drop the 10 seconds pause, the result is that 'a' is set to 101.<br/><br/>And in the practice timing changes! For instance the client may wait 10 seconds before INCR, but the sequence written in the Append Only File (and later replayed-back as fast as possible when Redis is restarted) will not have the pause. Even if we add a timestamp in the AOF, when the time difference is smaller than our timer resolution, we have a race condition.<br/><br/>The same happens with master-slave replication. Again, consider the example above: the client will use the same sequence of commands without the 10 seconds pause, but the replication link will slow down for a few seconds due to a network problem. Result? The master will contain 'a' set to 101, the slave 'a' set to 1.<br/><br/>The only way to avoid this but at the same time have reliable non time dependent timeouts on keys is to destroy volatile keys when a write operation is attempted against it.<br/><br/>After all Redis is one of the rare fully persistent databases that will give you EXPIRE. This comes to a cost :)<h2><a name="FAQ: How this limitations were solved in Redis versions > 2.1.3?">FAQ: How this limitations were solved in Redis versions > 2.1.3?</a></h2>Since Redis 2.1.3 there are no longer restrictions in the use you can do of write commands against volatile keys, still the replication and AOF file are guaranteed to be fully consistent.<br/><br/>In order to obtain a correct behavior without sacrificing consistency now when a key expires, a DEL operation is synthesized in both the AOF file and against all the attached slaves. This way the expiration process is centralized in the master instance, and there is no longer a chance of consistency errors.<br/><br/>However while the slaves while connected to a master will not expire keys independently, they'll still take the full state of the expires existing in the dataset, so when a slave is elected to a master it will be able to expire the keys independently, fully acting as a master.
projects / redis.git / blobdiff