A few updates to the thread. I uploaded a patch[1] as a complete example of how users can use the metrics in the future.
Some thoughts below after taking a look at the AbstractMetricGroup and its subclasses. This patch intends to provide convenience for Flink connector implementations to follow metrics standards proposed in FLIP-33. It also try to enhance the metric management in general way to help users with: 1. metric definition 2. metric dependencies check 3. metric validation 4. metric control (turn on / off particular metrics) This patch wraps MetricGroup to extend the functionality of AbstractMetricGroup and its subclasses. The AbstractMetricGroup mainly focus on the metric group hierarchy, but does not really manage the metrics other than keeping them in a Map. Ideally we should only have one entry point for the metrics. Right now the entry point is AbstractMetricGroup. However, besides the missing functionality mentioned above, AbstractMetricGroup seems deeply rooted in Flink runtime. We could extract it out to flink-metrics in order to use it for generic purpose. There will be some work, though. Another approach is to make AbstractMetrics in this patch as the metric entry point. It wraps metric group and provides the missing functionalities. Then we can roll out this pattern to runtime components gradually as well. My first thought is that the latter approach gives a more smooth migration. But I am also OK with doing a refactoring on the AbstractMetricGroup family. Thanks, Jiangjie (Becket) Qin [1] https://github.com/becketqin/flink/pull/1 On Mon, Feb 25, 2019 at 2:32 PM Becket Qin <becket....@gmail.com> wrote: > Hi Chesnay, > > It might be easier to discuss some implementation details in the PR review > instead of in the FLIP discussion thread. I have a patch for Kafka > connectors ready but haven't submitted the PR yet. Hopefully that will help > explain a bit more. > > ** Re: metric type binding > This is a valid point that worths discussing. If I understand correctly, > there are two points: > > 1. Metric type / interface does not matter as long as the metric semantic > is clearly defined. > Conceptually speaking, I agree that as long as the metric semantic is > defined, metric type does not matter. To some extent, Gauge / Counter / > Meter / Histogram themselves can be think of as some well-recognized > semantics, if you wish. In Flink, these metric semantics have their > associated interface classes. In practice, such semantic to interface > binding seems necessary for different components to communicate. Simply > standardize the semantic of the connector metrics seems not sufficient for > people to build ecosystem on top of. At the end of the day, we still need > to have some embodiment of the metric semantics that people can program > against. > > 2. Sometimes the same metric semantic can be exposed using different > metric types / interfaces. > This is a good point. Counter and Gauge-as-a-Counter are pretty much > interchangeable. This is more of a trade-off between the user experience of > metric producers and consumers. The metric producers want to use Counter or > Gauge depending on whether the counter is already tracked in code, while > ideally the metric consumers only want to see a single metric type for each > metric. I am leaning towards to make the metric producers happy, i.e. allow > Gauge / Counter metric type, and the the metric consumers handle the type > variation. The reason is that in practice, there might be more connector > implementations than metric reporter implementations. We could also provide > some helper method to facilitate reading from such variable metric type. > > > Just some quick replies to the comments around implementation details. > >> 4) single place where metrics are registered except connector-specific >> ones (which we can't really avoid). > > Register connector specific ones in a single place is actually something > that I want to achieve. > > 2) I'm talking about time-series databases like Prometheus. We would >> only have a gauge metric exposing the last fetchTime/emitTime that is >> regularly reported to the backend (Prometheus), where a user could build >> a histogram of his choosing when/if he wants it. >> > Not sure if such downsampling works. As an example, if a user complains > that there are some intermittent latency spikes (maybe a few records in 10 > seconds) in their processing system. Having a Gauge sampling instantaneous > latency seems unlikely useful. However by looking at actual 99.9 percentile > latency might help. > > Thanks, > > Jiangjie (Becket) Qin > > > On Fri, Feb 22, 2019 at 9:30 PM Chesnay Schepler <ches...@apache.org> > wrote: > >> Re: over complication of implementation. >> >> I think I get understand better know what you're shooting for, >> effectively something like the OperatorIOMetricGroup. >> But still, re-define setupConnectorMetrics() to accept a set of flags >> for counters/meters(ans _possibly_ histograms) along with a set of >> well-defined Optional<Gauge<?>>, and return the group. >> >> Solves all issues as far as i can tell: >> 1) no metrics must be created manually (except Gauges, which are >> effectively just Suppliers and you can't get around this), >> 2) additional metrics can be registered on the returned group, >> 3) see 1), >> 4) single place where metrics are registered except connector-specific >> ones (which we can't really avoid). >> >> Re: Histogram >> >> 1) As an example, whether "numRecordsIn" is exposed as a Counter or a >> Gauge should be irrelevant. So far we're using the metric type that is >> the most convenient at exposing a given value. If there is some backing >> data-structure that we want to expose some data from we typically opt >> for a Gauge, as otherwise we're just mucking around with the >> Meter/Counter API to get it to match. Similarly, if we want to count >> something but no current count exists we typically added a Counter. >> That's why attaching semantics to metric types makes little sense (but >> unfortunately several reporters already do it); for counters/meters >> certainly, but the majority of metrics are gauges. >> >> 2) I'm talking about time-series databases like Prometheus. We would >> only have a gauge metric exposing the last fetchTime/emitTime that is >> regularly reported to the backend (Prometheus), where a user could build >> a histogram of his choosing when/if he wants it. >> >> On 22.02.2019 13:57, Becket Qin wrote: >> > Hi Chesnay, >> > >> > Thanks for the explanation. >> > >> > ** Re: FLIP >> > I might have misunderstood this, but it seems that "major changes" are >> well >> > defined in FLIP. The full contents is following: >> > What is considered a "major change" that needs a FLIP? >> > >> > Any of the following should be considered a major change: >> > >> > - Any major new feature, subsystem, or piece of functionality >> > - *Any change that impacts the public interfaces of the project* >> > >> > What are the "public interfaces" of the project? >> > >> > >> > >> > *All of the following are public interfaces *that people build around: >> > >> > - DataStream and DataSet API, including classes related to that, >> such as >> > StreamExecutionEnvironment >> > >> > >> > - Classes marked with the @Public annotation >> > >> > >> > - On-disk binary formats, such as checkpoints/savepoints >> > >> > >> > - User-facing scripts/command-line tools, i.e. bin/flink, Yarn >> scripts, >> > Mesos scripts >> > >> > >> > - Configuration settings >> > >> > >> > - *Exposed monitoring information* >> > >> > >> > So any monitoring information change is considered as public interface, >> and >> > any public interface change is considered as a "major change". >> > >> > >> > ** Re: over complication of implementation. >> > >> > Although this is more of implementation details that is not covered by >> the >> > FLIP. But it may be worth discussing. >> > >> > First of all, I completely agree that we should use the simplest way to >> > achieve our goal. To me the goal is the following: >> > 1. Clear connector conventions and interfaces. >> > 2. The easiness of creating a connector. >> > >> > Both of them are important to the prosperity of the connector >> ecosystem. So >> > I'd rather abstract as much as possible on our side to make the >> connector >> > developer's work lighter. Given this goal, a static util method approach >> > might have a few drawbacks: >> > 1. Users still have to construct the metrics by themselves. (And note >> that >> > this might be erroneous by itself. For example, a customer wrapper >> around >> > dropwizard meter maybe used instead of MeterView). >> > 2. When connector specific metrics are added, it is difficult to enforce >> > the scope to be the same as standard metrics. >> > 3. It seems that a method proliferation is inevitable if we want to >> apply >> > sanity checks. e.g. The metric of numBytesIn was not registered for a >> meter. >> > 4. Metrics are still defined in random places and hard to track. >> > >> > The current PR I had was inspired by the Config system in Kafka, which I >> > found pretty handy. In fact it is not only used by Kafka itself but even >> > some other projects that depend on Kafka. I am not saying this approach >> is >> > perfect. But I think it worths to save the work for connector writers >> and >> > encourage more systematic implementation. That being said, I am fully >> open >> > to suggestions. >> > >> > >> > Re: Histogram >> > I think there are two orthogonal questions around those metrics: >> > >> > 1. Regardless of the metric type, by just looking at the meaning of a >> > metric, is generic to all connectors? If the answer is yes, we should >> > include the metric into the convention. No matter whether we include it >> > into the convention or not, some connector implementations will emit >> such >> > metric. It is better to have a convention than letting each connector do >> > random things. >> > >> > 2. If a standard metric is a histogram, what should we do? >> > I agree that we should make it clear that using histograms will have >> > performance risk. But I do see histogram is useful in some >> fine-granularity >> > debugging where one do not have the luxury to stop the system and inject >> > more inspection code. So the workaround I am thinking is to provide some >> > implementation suggestions. Assume later on we have a mechanism of >> > selective metrics. In the abstract metrics class we can disable those >> > metrics by default individual connector writers does not have to do >> > anything (this is another advantage of having an AbstractMetrics >> instead of >> > static util methods.) >> > >> > I am not sure I fully understand the histogram in the backend approach. >> Can >> > you explain a bit more? Do you mean emitting the raw data, e.g. >> fetchTime >> > and emitTime with each record and let the histogram computation happen >> in >> > the background? Or let the processing thread putting the values into a >> > queue and have a separate thread polling from the queue and add them >> into >> > the histogram? >> > >> > Thanks, >> > >> > Jiangjie (Becket) Qin >> > >> > >> > >> > >> > >> > On Fri, Feb 22, 2019 at 4:34 PM Chesnay Schepler <ches...@apache.org> >> wrote: >> > >> >> Re: Flip >> >> The very first line under both the main header and Purpose section >> >> describe Flips as "major changes", which this isn't. >> >> >> >> Re: complication >> >> I'm not arguing against standardization, but again an over-complicated >> >> implementation when a static utility method would be sufficient. >> >> >> >> public static void setupConnectorMetrics( >> >> MetricGroup operatorMetricGroup, >> >> String connectorName, >> >> Optional<Gauge<Long>> numRecordsIn, >> >> ...) >> >> >> >> This gives you all you need: >> >> * a well-defined set of metrics for a connector to opt-in >> >> * standardized naming schemes for scope and individual metrics >> >> * standardize metric types (although personally I'm not interested in >> that >> >> since metric types should be considered syntactic sugar) >> >> >> >> Re: Configurable Histogram >> >> If anything they _must_ be turned off by default, but the metric >> system is >> >> already exposing so many options that I'm not too keen on adding even >> more. >> >> You have also only addressed my first argument against histograms >> >> (performance), the second one still stands (calculate histogram in >> metric >> >> backends instead). >> >> >> >> On 21.02.2019 16:27, Becket Qin wrote: >> >>> Hi Chesnay, >> >>> >> >>> Thanks for the comments. I think this is worthy of a FLIP because it >> is >> >>> public API. According to the FLIP description a FlIP is required in >> case >> >> of: >> >>> - Any change that impacts the public interfaces of the project >> >>> >> >>> and the following entry is found in the definition of "public >> interface". >> >>> >> >>> - Exposed monitoring information >> >>> >> >>> Metrics are critical to any production system. So a clear metric >> >> definition >> >>> is important for any serious users. For an organization with large >> Flink >> >>> installation, change in metrics means great amount of work. So such >> >> changes >> >>> do need to be fully discussed and documented. >> >>> >> >>> ** Re: Histogram. >> >>> We can discuss whether there is a better way to expose metrics that >> are >> >>> suitable for histograms. My micro-benchmark on various histogram >> >>> implementations also indicates that they are significantly slower than >> >>> other metric types. But I don't think that means never use histogram, >> but >> >>> means use it with caution. For example, we can suggest the >> >> implementations >> >>> to turn them off by default and only turn it on for a small amount of >> >> time >> >>> when performing some micro-debugging. >> >>> >> >>> ** Re: complication: >> >>> Connector conventions are essential for Flink ecosystem. Flink >> connectors >> >>> pool is probably the most important part of Flink, just like any other >> >> data >> >>> system. Clear conventions of connectors will help build Flink >> ecosystem >> >> in >> >>> a more organic way. >> >>> Take the metrics convention as an example, imagine someone has >> developed >> >> a >> >>> Flink connector for System foo, and another developer may have >> developed >> >> a >> >>> monitoring and diagnostic framework for Flink which analyzes the Flink >> >> job >> >>> performance based on metrics. With a clear metric convention, those >> two >> >>> projects could be developed independently. Once users put them >> together, >> >>> it would work without additional modifications. This cannot be easily >> >>> achieved by just defining a few constants. >> >>> >> >>> ** Re: selective metrics: >> >>> Sure, we can discuss that in a separate thread. >> >>> >> >>> @Dawid >> >>> >> >>> ** Re: latency / fetchedLatency >> >>> The primary purpose of establish such a convention is to help >> developers >> >>> write connectors in a more compatible way. The convention is supposed >> to >> >> be >> >>> defined more proactively. So when look at the convention, it seems >> more >> >>> important to see if the concept is applicable to connectors in >> general. >> >> It >> >>> might be true so far only Kafka connector reports latency. But there >> >> might >> >>> be hundreds of other connector implementations in the Flink ecosystem, >> >>> though not in the Flink repo, and some of them also emits latency. I >> >> think >> >>> a lot of other sources actually also has an append timestamp. e.g. >> >> database >> >>> bin logs and some K-V stores. So I wouldn't be surprised if some >> database >> >>> connector can also emit latency metrics. >> >>> >> >>> Thanks, >> >>> >> >>> Jiangjie (Becket) Qin >> >>> >> >>> >> >>> On Thu, Feb 21, 2019 at 10:14 PM Chesnay Schepler <ches...@apache.org >> > >> >>> wrote: >> >>> >> >>>> Regarding 2) It doesn't make sense to investigate this as part of >> this >> >>>> FLIP. This is something that could be of interest for the entire >> metric >> >>>> system, and should be designed for as such. >> >>>> >> >>>> Regarding the proposal as a whole: >> >>>> >> >>>> Histogram metrics shall not be added to the core of Flink. They are >> >>>> significantly more expensive than other metrics, and calculating >> >>>> histograms in the application is regarded as an anti-pattern by >> several >> >>>> metric backends, who instead recommend to expose the raw data and >> >>>> calculate the histogram in the backend. >> >>>> >> >>>> Second, this seems overly complicated. Given that we already >> established >> >>>> that not all connectors will export all metrics we are effectively >> >>>> reducing this down to a consistent naming scheme. We don't need >> anything >> >>>> sophisticated for that; basically just a few constants that all >> >>>> connectors use. >> >>>> >> >>>> I'm not convinced that this is worthy of a FLIP. >> >>>> >> >>>> On 21.02.2019 14:26, Dawid Wysakowicz wrote: >> >>>>> Hi, >> >>>>> >> >>>>> Ad 1. In general I undestand and I agree. But those particular >> metrics >> >>>>> (latency, fetchLatency), right now would only be reported if user >> uses >> >>>>> KafkaConsumer with internal timestampAssigner with >> StreamCharacteristic >> >>>>> set to EventTime, right? That sounds like a very specific case. I am >> >> not >> >>>>> sure if we should introduce a generic metric that will be >> >>>>> disabled/absent for most of implementations. >> >>>>> >> >>>>> Ad.2 That sounds like an orthogonal issue, that might make sense to >> >>>>> investigate in the future. >> >>>>> >> >>>>> Best, >> >>>>> >> >>>>> Dawid >> >>>>> >> >>>>> On 21/02/2019 13:20, Becket Qin wrote: >> >>>>>> Hi Dawid, >> >>>>>> >> >>>>>> Thanks for the feedback. That makes sense to me. There are two >> cases >> >> to >> >>>> be >> >>>>>> addressed. >> >>>>>> >> >>>>>> 1. The metrics are supposed to be a guidance. It is likely that a >> >>>> connector >> >>>>>> only supports some but not all of the metrics. In that case, each >> >>>> connector >> >>>>>> implementation should have the freedom to decide which metrics are >> >>>>>> reported. For the metrics that are supported, the guidance should >> be >> >>>>>> followed. >> >>>>>> >> >>>>>> 2. Sometimes users may want to disable certain metrics for some >> reason >> >>>>>> (e.g. performance / reprocessing of data). A generic mechanism >> should >> >> be >> >>>>>> provided to allow user choose which metrics are reported. This >> >> mechanism >> >>>>>> should also be honored by the connector implementations. >> >>>>>> >> >>>>>> Does this sound reasonable to you? >> >>>>>> >> >>>>>> Thanks, >> >>>>>> >> >>>>>> Jiangjie (Becket) Qin >> >>>>>> >> >>>>>> >> >>>>>> >> >>>>>> On Thu, Feb 21, 2019 at 4:22 PM Dawid Wysakowicz < >> >>>> dwysakow...@apache.org> >> >>>>>> wrote: >> >>>>>> >> >>>>>>> Hi, >> >>>>>>> >> >>>>>>> Generally I like the idea of having a unified, standard set of >> >> metrics >> >>>> for >> >>>>>>> all connectors. I have some slight concerns about fetchLatency and >> >>>>>>> latency though. They are computed based on EventTime which is not >> a >> >>>> purely >> >>>>>>> technical feature. It depends often on some business logic, might >> be >> >>>> absent >> >>>>>>> or defined after source. Those metrics could also behave in a >> weird >> >>>> way in >> >>>>>>> case of replaying backlog. Therefore I am not sure if we should >> >> include >> >>>>>>> those metrics by default. Maybe we could at least introduce a >> feature >> >>>>>>> switch for them? What do you think? >> >>>>>>> >> >>>>>>> Best, >> >>>>>>> >> >>>>>>> Dawid >> >>>>>>> On 21/02/2019 03:13, Becket Qin wrote: >> >>>>>>> >> >>>>>>> Bump. If there is no objections to the proposed metrics. I'll >> start a >> >>>>>>> voting thread later toady. >> >>>>>>> >> >>>>>>> Thanks, >> >>>>>>> >> >>>>>>> Jiangjie (Becket) Qin >> >>>>>>> >> >>>>>>> On Mon, Feb 11, 2019 at 8:17 PM Becket Qin <becket....@gmail.com> >> < >> >>>> becket....@gmail.com> wrote: >> >>>>>>> Hi folks, >> >>>>>>> >> >>>>>>> I would like to start the FLIP discussion thread about standardize >> >> the >> >>>>>>> connector metrics. >> >>>>>>> >> >>>>>>> In short, we would like to provide a convention of Flink connector >> >>>>>>> metrics. It will help simplify the monitoring and alerting on >> Flink >> >>>> jobs. >> >>>>>>> The FLIP link is following: >> >>>>>>> >> >>>>>>> >> >> >> https://cwiki.apache.org/confluence/display/FLINK/FLIP-33%3A+Standardize+Connector+Metrics >> >>>>>>> Thanks, >> >>>>>>> >> >>>>>>> Jiangjie (Becket) Qin >> >>>>>>> >> >>>>>>> >> >>>>>>> >> >> >> >>
