I was thinking on a smaller scale in terms of effort, but the more I think about it, the more supportive I would be of a full revamp (new API) for metrics based on Coda Hale's metrics library. It's proven and stable. I've used it many times. I think either approach would be roughly the same amount of work.
Some of the metrics API improvements in the 1.1.x branch are nice, but IMHO are lipstick on a pig. With apologies to Jungtaek, who has done amazing work all across the codebase, I'm a little squeamish about the proposed change to metrics that changes the consumer API based on a configuration flag (don't know the PR number offhand). I'm +1 for moving in this direction (revamped metrics). Let's end the metrics pain. -Taylor > On Oct 11, 2016, at 10:15 AM, Bobby Evans <ev...@yahoo-inc.com.INVALID> wrote: > > I agree that IMetricsConsumer is not good, but the reality is that all of the > metrics system needs to be redone. The problem is that we ship an object as > a metric. If I get an object I have no idea what it is hand hence no idea > how to report it or what to do with it. What is more the common types we use > in the metrics we provide are really not enough. For example CountMetric > sends a Long. Well when I get it in the metrics consumer I have no idea if I > should report it like a counter or if I should report it like a gauge > (something that every metrics system I have used wants to know). But then we > support pre-aggregation of the metrics with IReducer so the number I get > might be an average instead of either a gauge or a counter, which no good > metrics system will want to collect because I cannot aggregate it with > anything else, the math just does not work. > The proposal I have said before and I still believe is that we need to put in > place a parallel metrics API/system. We will deprecate all of > https://git.corp.yahoo.com/storm/storm/tree/master-security/storm-core/src/jvm/backtype/storm/metric/api > and create a new parallel one that provides an API similar to > http://metrics.dropwizard.io/3.1.0/. I would even be fine in just using > their API and exposing that to end users. Dropwizard has solved all of these > problems already and I don't see a reason to reinvent the wheel. I don't > personally see a lot of value in trying to send all of the metrics through > storm itslef. I am fine if we are able to support that, but it is far from a > requirement. - Bobby > > On Monday, October 10, 2016 10:47 PM, S G <sg.online.em...@gmail.com> > wrote: > > > +1 > We can probably start by opening a JIRA for this and adding a design > approach for the same? > I would like to help in the coding-effort for this. > > -SG > >> On Mon, Oct 10, 2016 at 1:51 PM, P. Taylor Goetz <ptgo...@gmail.com> wrote: >> >> I’ve been thinking about metrics lately, specifically the fact that people >> tend to struggle with implementing a metrics consumer. (Like this one [1]). >> >> The IMetricsConsumer interface is pretty low level, and common >> aggregations, calculations, etc. are left up to each individual >> implementation. That seems like an area where further abstraction would >> make it easier to support different back ends (Graphite, JMX, Splunk, etc.). >> >> My thought is to create an abstract IMetricsConsumer implementation that >> does common aggregations and calculations, and then delegates to a plugable >> “metrics sink” implementation (e.g. “IMetricsSink”, etc.). That would >> greatly simplify the effort required to integrate with various external >> metrics systems. I know of at least a few users that would be interested, >> one is currently scraping the logs from LoggingMetricsConsumer and polling >> the Storm REST API for their metrics. >> >> -Taylor >> >> [1] http://twocentsonsoftware.blogspot.co.il/2014/12/ >> sending-out-storm-metrics.html >> >> >>> On Oct 10, 2016, at 12:14 PM, Bobby Evans <ev...@yahoo-inc.com.INVALID> >> wrote: >>> >>> First of all the server exposes essentially the same interface that the >> IMetricsConsumer exposes. It mostly just adds a bunch of overhead in the >> middle to serialize out the objects send them over http to another process >> which then has to deserialize them and process them. If you really don't >> need the metrics to show up on a special known box you can have that exact >> same code running inside the metrics consumer without all of the overhead. >>> The server/client are insecure, have to deal with thread issues that a >> normal IMetricsConsumer does not, and are not written to be robust (If the >> HTTP server is down the consumer crashes and continues to crash until the >> server is brought back up). It was written very quickly for a test >> situation and it honestly never crossed my mind that anyone would want to >> use it in production. >>> >>> - Bobby >>> >>> On Monday, October 10, 2016 10:59 AM, S G <sg.online.em...@gmail.com> >> wrote: >>> >>> >>> Thanks Bobby. >>> >>> If we write our own metrics consumer, how do we ensure that it is better >>> than HttpForwardingMetricsServer? In other words, what aspects of the >>> HttpForwardingMetricsServer >>> should we avoid to make our own metrics consumer better and ready for >>> production? >>> >>> Is versign/storm-graphite <https://github.com/verisign/storm-graphite> >>> production >>> ready? >>> >>> Also, we should add a line about production-readiness of >>> HttpForwardingMetricsServer >>> in the documentation at http://storm.apache.org/ >> releases/1.0.2/Metrics.html >>> (We were just about to think seriously on using this for production as we >>> thought this to be the standard solution for metrics' consumption in 1.0+ >>> version). >>> >>> -SG >>> >>> On Mon, Oct 10, 2016 at 6:37 AM, Bobby Evans <ev...@yahoo-inc.com> >> wrote: >>> >>>> First of all there really are two different sets of metrics. One set is >>>> the topology metrics and the other set is the daemon metrics (metrics >> for >>>> things like the ui and nimbus). The JmxPreparableReporter plugin only >>>> exposes daemon metrics not the topology metrics through JMX. Exposing >>>> topology metrics through JMX is a non trivial task. The current metrics >>>> feature was not designed for this. We are in the process of trying to >>>> redesign the metrics system to allow for features like this, but it is >>>> still a ways off. >>>> >>>> - Bobby >>>> >>>> >>>> On Saturday, October 8, 2016 11:39 AM, S G <sg.online.em...@gmail.com> >>>> wrote: >>>> >>>> >>>> Thanks Bobby, >>>> >>>> We will need some kind of IMetricsConsumer to talk to telegraf. >>>> Many other softwares like Solr, Elastic-Search, Cassandra etc. provide >>>> metrics through a URL making it very easy to consume by tools like >> telegraf. >>>> How about a IMetricsConsumer that will run on storm-ui and provide the >>>> metrics through a URL such as <storm-ui-host>/metrics ? >>>> >>>> Also, I see the following option in defaults.yaml: >>>> #default storm daemon metrics reporter plugins >>>> storm.daemon.metrics.reporter.plugins: >>>> - "org.apache.storm.daemon.metrics.reporters. >> JmxPreparableReporter" >>>> >>>> Is this a good option to use for converting metrics into JMX ? >>>> >>>> Thanks >>>> SG >>>> >>>> >>>> >>>> >>>> On Fri, Oct 7, 2016 at 8:11 AM, Bobby Evans <ev...@yahoo-inc.com.invalid >>> >>>> wrote: >>>> >>>> HttpForwardingMetricsServer is a real hack intended really for tests. I >>>> know I wrote it :). Please don't use it in production. You can write >> your >>>> own IMetricesConsumer to do whatever you want to with the metrics. >>>> https://github.com/apache/ storm/blob/master/storm-core/ >>>> src/jvm/org/apache/storm/ metric/api/IMetricsConsumer. java >>>> <https://github.com/apache/storm/blob/master/storm-core/ >> src/jvm/org/apache/storm/metric/api/IMetricsConsumer.java> >>>> >>>> That is the correct way to get the data out. If you want to write a >>>> bridge to JMX for this that might work, but going directly to telegraph >>>> would probably be better. - Bobby >>>> >>>> On Thursday, October 6, 2016 1:43 PM, S G < >> sg.online.em...@gmail.com> >>>> wrote: >>>> >>>> >>>> Hi, >>>> >>>> We want to use Telegraf ( >>>> https://github.com/influxdata/ telegraf/tree/master/plugins >>>> <https://github.com/influxdata/telegraf/tree/master/plugins>) for >> getting >>>> storm's metrics. >>>> >>>> But we do not want to add a HttpForwardingMetricsServer just to get the >>>> metrics and send them to telegraf. >>>> >>>> Other option is to use Jolokia (https://jolokia.org/) that can read JMX >>>> and >>>> write into telegraf. >>>> >>>> Does storm report all its metrics (including those of custom >> spouts/bolts) >>>> into JMX? >>>> Or spawning a HttpForwardingMetricsServer is the only option? >>>> >>>> Thanks >>>> SG >>>> >>>> >>>> >>>> >>>> >>>> >>>> >>>> >>> >>> >> >> >
