Agreed…a global knob is a bit tricky for exactly the reason you've identified. Perhaps the problem could be simplified though by considering the context and purpose of Kafka. I would use a persistent message queue because I want to guarantee that data/messages don't get lost. But, since Kafka is not meant to be a long term storage solution (other products can be used for that), I would clarify that guarantee to apply only to the most recent messages up until a certain configured threshold (i.e. max 24 hrs, max 500GB, etc). Once those thresholds are reached, old messages are deleted first.
To ensure no message loss (up to a limit), I must ensure Kafka is highly available. There's a small a chance that the message deletion rate is the same rate that receive rate. For example, when the incoming volume is so high that the size threshold is reached before the time threshold. But, I may be ok with that because if Kafka goes down, it can cause upstream applications to fail. This can result in higher losses overall, and particularly of the most *recent* messages. In other words, in a persistent but ephemeral message queue, I would give higher precedence to recent messages over older ones. On the flip side, by allowing Kafka to go down when a disk is full, applications are forced to deal with the issue. This adds complexity to apps, but perhaps it's not a bad thing. After all, in scalability, all apps should be designed to handle failure. Having said that, next is to decide which messages to delete first. I believe that's a separate issue and has its own complexities, too. The main idea though is that a global knob would provide flexibility, even if not used. From an operation perspective, if we can't ensure HA for all applications/components, it would be good if we can for at least some of the core ones, like Kafka. This is much easier said that done though. On May 5, 2014, at 9:16 AM, Jun Rao <jun...@gmail.com> wrote: > Yes, your understanding is correct. A global knob that controls aggregate > log size may make sense. What would be the expected behavior when that > limit is reached? Would you reduce the retention uniformly across all > topics? Then, it just means that some of the logs may not be retained as > long as you want. Also, we need to think through what happens when every > log has only 1 segment left and yet the total size still exceeds the limit. > Do we roll log segments early? > > Thanks, > > Jun > > > On Sun, May 4, 2014 at 4:31 AM, vinh <v...@loggly.com> wrote: > >> Thanks Jun. So if I understand this correctly, there really is no master >> property to control the total aggregate size of all Kafka data files on a >> broker. >> >> log.retention.size and log.file.size are great for managing data at the >> application level. In our case, application needs change frequently, and >> performance itself is an ever evolving feature. This means various configs >> are constantly changing, like topics, # of partitions, etc. >> >> What rarely changes though is provisioned hardware resources. So a >> setting to control the total aggregate size of Kafka logs (or persisted >> data, for better clarity) would definitely simplify things at an >> operational level, regardless what happens at the application level. >> >> >> On May 2, 2014, at 7:49 AM, Jun Rao <jun...@gmail.com> wrote: >> >>> log.retention.size controls the total size in a log dir (per >>> partition). log.file.size >>> controls the size of each log segment in the log dir. >>> >>> Thanks, >>> >>> Jun >>> >>> >>> On Thu, May 1, 2014 at 9:31 PM, vinh <v...@loggly.com> wrote: >>> >>>> In the 0.7 docs, the description for log.retention.size and >> log.file.size >>>> sound very much the same. In particular, that they apply to a single >> log >>>> file (or log segment file). >>>> >>>> http://kafka.apache.org/07/configuration.html >>>> >>>> I'm beginning to think there is no setting to control the max aggregate >>>> size of all logs. If this is correct, what would be a good approach to >>>> enforce this requirement? In my particular scenario, I have a lot of >> data >>>> being written to Kafka at a very high rate. So a 1TB disk can easily be >>>> filled up in 24hrs or so. One option is to add more Kafka brokers to >> add >>>> more disk space to the pool, but I'd like to avoid that and see if I can >>>> simply configure Kafka to not write more than 1TB aggregate. Else, >> Kafka >>>> will OOM and kill itself, and possibly the crash the node itself because >>>> the disk is full. >>>> >>>> >>>> On May 1, 2014, at 9:21 PM, vinh <v...@loggly.com> wrote: >>>> >>>>> Using Kafka 0.7.2, I have the following in server.properties: >>>>> >>>>> log.retention.hours=48 >>>>> log.retention.size=107374182400 >>>>> log.file.size=536870912 >>>>> >>>>> My interpretation of this is: >>>>> a) a single log segment file over 48hrs old will be deleted >>>>> b) the total combined size of *all* logs is 100GB >>>>> c) a single log segment file is limited to 500MB in size before a new >>>> segment file is spawned spawning a new segment file >>>>> d) a "log file" can be composed of many "log segment files" >>>>> >>>>> But, even after setting the above, I find that the total combined size >>>> of all Kafka logs on disk is 200GB right now. Isn't log.retention.size >>>> supposed to limit it to 100GB? Am I missing something? The docs are >> not >>>> really clear, especially when it comes to distinguishing between a "log >>>> file" and a "log segment file". >>>>> >>>>> I have disk monitoring. But like anything else in software, even >>>> monitoring can fail. Via configuration, I'd like to make sure that >> Kafka >>>> does not write more than the available disk space. Or something like >>>> log4j, where I can set a max number of log files and the max size per >> file, >>>> which essentially allows me to set a max aggregate size limit across all >>>> logs. >>>>> >>>>> Thanks, >>>>> -Vinh >>>> >>>> >> >>
