For the samples: it will be at least double that estimate I think since the long array contains (eight byte) references to the actual longs, each of which also have some object overhead.
Re: testing: actually, it looks like YM metrics does allow you to drop in your own clock: https://github.com/dropwizard/metrics/blob/master/metrics-core/src/main/java/com/codahale/metrics/Clock.java https://github.com/dropwizard/metrics/blob/master/metrics-core/src/main/java/com/codahale/metrics/Meter.java#L36 Not sure if it was mentioned in this (or some recent) thread but a major motivation in the kafka-common metrics (KM) was absorbing API changes and even mbean naming conventions. For e.g., in the early stages of 0.8 we picked up YM metrics 3.x but collided with client apps at LinkedIn which were still on 2.x. We ended up changing our code to use 2.x in the end. Having our own metrics package makes us less vulnerable to these kinds of changes. The multiple version collision problem is obviously less of an issue with the broker but we are still exposed to possible metric changes in YM metrics. I'm wondering if we need to weigh too much toward the memory overheads of histograms in making a decision here simply because I don't think we have found them to be an extreme necessity for per-clientid/per-partition metrics and they are more critical for aggregate (global) metrics. So it seems the main benefits of switching to KM metrics are: - Less exposure to YM metrics changes - More control over the actual implementation. E.g., there is considerable research on implementing approximate-but-good-enough histograms/percentiles that we can try out - Differences (improvements) from YM metrics such as: - hierarchical sensors - integrated with quota enforcement - mbeans can logically group attributes computed from different sensors. So there is logical grouping (as opposed to a separate mbean per sensor as is the case in YM metrics). The main disadvantages: - Everyone's graphs and alerts will break and need to be updated - Histogram support needs to be tested more/improved The first disadvantage is a big one but we aren't exactly immune to that if we stick with YM. BTW with KM metrics we should also provide reporters (graphite, ganglia) but we probably need to do this anyway since the new clients are on KM metrics. Thanks, Joel On Fri, Mar 27, 2015 at 06:48:48PM +0000, Aditya Auradkar wrote: > Adding to what Jay said. > > The library maintains 1k samples by default. The UniformSample has a long > array so about 8k overhead per histogram. The ExponentiallyDecayingSample > (which is what we use) has a 16 byte overhead per stored sample, so about 16k > per histogram. So 10k histograms (worst case? metrics per partition and > client) is about 160MB of memory in the broker. > > Copying is also a problem. For percentiles on HistogramMBean, the > implementation does a copy of the entire array. For e.g., if we called > get50Percentile() and get75Percentile(), the entire array would get copied > twice which is pretty bad if we called each metric on every MBean. > > Another point Joel mentioned is that codahale metrics are harder to write > tests against because we cannot pass in a Clock. > > IMO, if a library is preventing us from adding all the metrics that we want > to add and we have a viable alternative, we should replace it. It might be > short term pain but in the long run we will have more useful graphs. > What do people think? I can start a vote thread on this once we have a couple > more opinions. > > Thanks, > Aditya > ________________________________________ > From: Jay Kreps [jay.kr...@gmail.com] > Sent: Thursday, March 26, 2015 2:29 PM > To: dev@kafka.apache.org > Subject: Re: Metrics package discussion > > Yeah that is a good summary. > > The reason we don't use histograms heavily in the server is because of the > memory issues. We originally did use histograms for everything, then we ran > into all these issues, and ripped them out. Whether they are really useful > or not, I don't know. Averages can be pretty misleading so it can be nice > but I don't know that it is critical. > > -Jay > > On Thu, Mar 26, 2015 at 1:58 PM, Aditya Auradkar < > aaurad...@linkedin.com.invalid> wrote: > > > From what I can tell, Histograms don't seem to be used extensively in the > > Kafka server (only in RequestChannel.scala) and I'm not sure we need them > > for per-client metrics. Topic metrics use meters currently. Migrating > > graphing, alerting will be quite a significant effort for all users of > > Kafka. Do the potential benefits of the new metrics package outweigh this > > one time migration? In the long run it seems nice to have a unified metrics > > package across clients and server. If we were starting out from scratch > > without any existing deployments, what decision would we take? > > > > I suppose the relative effort in supporting is a useful data point in this > > discussion. We need to throttle based on the current byte rate which should > > be a "Meter" in codahale terms. The Meter implementation uses a 1, 5 and 15 > > minute exponential window moving average. The library also does not use the > > most recent samples of data for Metered metrics. For calculating rates, the > > EWMA class has a scheduled task that runs every 5 seconds and adjusts the > > rate using the new data accordingly. In that particular case, I think the > > new library is superior since it is more responsive. If we do choose to > > remain with Yammer on the server, here are a few ideas on how to support > > quotas with relatively less effort. > > > > - We could have a new type of Meter called "QuotaMeter" that can wrap the > > existing meter code that follows the same pattern that the Sensor does in > > the new metrics library. This QuotaMeter needs to be configured with a > > Quota and it can have a finer grained rate than 1 minute (10 seconds? > > configurable?). Anytime we call "mark()", it update the underlying rates > > and throw a QuotaViolationException if required. This class can either > > extend Meter or be a separate implementation of the Metric superclass that > > every metric implements. > > > > - We can also consider implementing these quotas with the new metrics > > package and have these co-exist with the existing metrics. This leads to 2 > > metric packages being used on the server, but they are both pulled in as > > dependencies anyway. Using this for metrics we can quota on may not be a > > bad place to start. > > > > Thanks, > > Aditya > > ________________________________________ > > From: Jay Kreps [jay.kr...@gmail.com] > > Sent: Wednesday, March 25, 2015 11:08 PM > > To: dev@kafka.apache.org > > Subject: Re: Metrics package discussion > > > > Here was my understanding of the issue last time. > > > > The yammer metrics use a random sample of requests to estimate the > > histogram. This allocates a fairly large array of longs (their values are > > longs rather than floats). A reasonable sample might be 8k entries which > > would give about 64KB per histogram. There are bounds on accuracy, but they > > are only probabilistic. I.e. if you try to get 99% < 5 ms of inaccuracy, > > you will 1% of the time get more than this. This is okay but if you try to > > alert, in which you realize that being wrong 1% of the time is a lot if you > > are computing stats every second continuously on many metrics (i.e. 1 in > > 100 estimates will be outside you bound). This array is copied in full > > every time you check the metric which is the other cause of the memory > > pressure. > > > > The better approach to histograms is to calculate buckets boundaries and > > record arbitrarily many values in those buckets. A simple bucketing > > approach for latency would be 0, 5ms, 10ms, 15ms, etc, and you just count > > how many fall in each bucket. Your precision is deterministically bounded > > by the bucket boundaries, so if you had 5ms buckets you would never have > > more than 5ms loss of precision. By using non-uniform bucket sizes you can > > make this work even better (e.g. give ~1ms precision for latencies in the > > 1ms range, but give only 1 second precision for latencies in the 30 second > > range). That is what is implemented in that metrics package. > > > > I think this bucketing approach is popular now. There is a whole "HDR > > histogram" library that gives lots of different bucketing methods and > > implements dynamic resizing so you don't have to specify an upper bound. > > https://github.com/HdrHistogram/HdrHistogram > > > > Whether this matters depends entirely if you want histograms broken down at > > the client, topic, partition, or broker level or just want overall metrics. > > If we just want per sever aggregates for histograms then I think the memory > > usage is not a huge issue. If you want a histogram per topic or client or > > partition and have 10k of these then that is where you start talking like > > 1GB of memory with the yammer package, which is what we hit last time. > > Getting percentiles on the client level is nice, percentiles are definitely > > better than averages, but I'm not sure it is required. > > > > -Jay > > > > On Wed, Mar 25, 2015 at 9:43 PM, Neha Narkhede <n...@confluent.io> wrote: > > > > > Aditya, > > > > > > If we are doing a deep dive, one of the things to investigate would be > > > memory/GC performance. IIRC, when I was looking into codahale at > > LinkedIn, > > > I remember it having quite a few memory management and GC issues while > > > using histograms. In comparison, histograms in the new metrics package > > > aren't very well tested. > > > > > > Thanks, > > > Neha > > > > > > On Wed, Mar 25, 2015 at 8:25 AM, Aditya Auradkar < > > > aaurad...@linkedin.com.invalid> wrote: > > > > > > > Hey everyone, > > > > > > > > Picking up this discussion after yesterdays KIP hangout. For anyone who > > > > did not join the meeting, we have 2 different metrics packages being > > used > > > > by the clients (custom package) and the server (codahale). We are > > > > discussing whether to migrate the server to the new package. > > > > > > > > What information do we need in order to make a decision? > > > > > > > > Some pros of the new package: > > > > - Using the most recent information by combining data from previous and > > > > current samples. I'm not sure how codahale does this so I'll > > investigate. > > > > - We can quota on anything we measure. This is pretty cool IMO. I've > > > > investigate the feasibility of adding this feature in codahale. > > > > - Hierarchical metrics. For example: we can define a sensor for overall > > > > bytes-in/bytes-out and also per-client. Updating the client sensor will > > > > cause the global byte rate sensor to get modified too. > > > > > > > > What are some of the issues with codahale? One previous discussion > > > > mentions high memory usage but I don't have any experience with it > > > myself. > > > > > > > > Thanks, > > > > Aditya > > > > > > > > > > > > > > > > > > > > > > > > > > > > > -- > > > Thanks, > > > Neha > > > > >
