Pretty slick. A few questions: * Can you use Lua scripts with blocking operations like BRPOP? Or is the strategy here to just poll if the queue is empty? (Eith

Pretty slick. A few questions: * Can you use Lua scripts with blocking operations like BRPOP? Or is the strategy here to just poll if the queue is empty? (Either is a usable strategy.) * Sometimes you will have particular queue items that cause the processor to die, because of some weird bug. It can be helpful to stick those into a separate "holding area" after a few failures, and send out warning emails to someone who can fix the problem. * How do you monitor the number of jobs queued and in progress? You could get all the items in the list with LRANGE and look at their timestamps (if any), but if the items are large and there are a few million of them, this could be irritatingly slow. Monitoring is super-important! That said, Redis can be really useful for writing queues. At Greplin we're pushing so much data through a Redis-based queue system that we actually managed to max out the operations-per-second on EC2 Large instances, and had to shard across a few servers. The basic idea is that we encode everything as protocol buffers, LPUSH it into a waiting queue, and have the processor do something similar to this: 1. RPOPLPUSH moves the item at the head of the waiting queue into a one-element in-progress queue. The name of the queue includes a counter for how many unsuccessful processing attempts have been made (starting at 0). 2. Process the item. 3. LPUSH the output into zero or more waiting queues for the next processing stage, and delete the in-progress queue. Unless there was an error, in which case things get complicated. :-( There are a surprising number of complicated little issues that come up when making queueing systems for heavy production use! For example, how do we coordinate the topology, and which servers to connect to? If one of our EC2 instances goes down and takes a Redis master with it, how do we seamlessly switch over to the slave? How do we shunt aside "killer" items for later processing after we've fixed the bugs that they were causing? What happens when there's a load spike and queue length goes through the roof, and the Redis servers start using up all their memory? These problems all have solutions -- for example, you can build most of the coordination and rate-limiting stuff with Zookeeper -- and for the most part I've been very happy with Redis. It's been very solid and reliable, with convenient operations, easy administration, and great speed! I just think that queueing systems are an area ripe for improvement.