I definitely agree that you need this level of understanding of your cluster. It definitely could work the way that you describe.
I was thinking of it slightly differently though. The metrics for this purpose (understanding performance of existing cluster) should come from the actual sensors themselves. For example, I need to instrument the packet capture process so that it kicks out time-series-ish metrics that you can monitor in a dashboard over time. On Fri, Apr 15, 2016 at 1:40 PM, [email protected] <[email protected]> wrote: > However, it would be handy to have something like this perpetually running > so you know when to scale up/out/down/in a cluster. > > On Fri, Apr 15, 2016, 13:35 Nick Allen <[email protected]> wrote: > > > I think it is slightly different. I don't even want to install minimal > > Kafka infrastructure (Look ma, no Kafka!) > > > > The exact implementation would differ based on the data inputs that you > are > > trying to measure, but for example... > > > > - To understand raw packet rates I would have a specialized sensor > that > > counts packets and size on the wire. It doesn't do anything more than > > that. > > - To understand Netflow rates, it would watch for Netflow packets and > > count those. > > - To understand sizing around application logs, a sensor would watch > for > > Syslog packets and count those. > > > > The implementation would be more similar to raw packet capture with some > > DPI. No Hadoop-y components required. > > > > > > > > On Fri, Apr 15, 2016 at 1:10 PM, James Sirota <[email protected]> > > wrote: > > > > > So this is exactly what I am proposing. Calculate the metrics on the > fly > > > without landing any data in the cluster. The problem is that that > > > enterprise data volumes are so large you can’t just point them at a > Java > > or > > > a C++ program or sensor. You either need an existing minimal Kafka > > > infrastructure to take that load or sample the data. > > > > > > Thanks, > > > James > > > > > > > > > > > > > > > On 4/15/16, 9:54 AM, "Nick Allen" <[email protected]> wrote: > > > > > > >Or we have the assessment tool not actually land any data. The > > assessment > > > >tool becomes a 'sensor' in its own right. You just point the input > data > > > >sets at the assessment tool, it builds metrics on the input (for > > example: > > > >count the number of packets per second) and then we use those metrics > to > > > >estimate cluster size. > > > > > > > >On Wed, Apr 13, 2016 at 5:45 PM, James Sirota < > [email protected]> > > > >wrote: > > > > > > > >> That’s an excellent point. So I think there are three ways forward. > > > >> > > > >> One is we can assume that there has to be at least a minimal > > > >> infrastructure in place (at least a subset of Kafka and Storm > > > resources) to > > > >> run a full-scale assessment. If you point something that blasts > > > millions > > > >> of messages per second at something like ActiveMQ you are going to > > blow > > > >> up. So the infrastructure to at least receive these kinds of > message > > > >> volumes has to exist as a pre-requisite. There is no way to get > around > > > that. > > > >> > > > >> The second approach I see is sampling. Sampling is a lot less > precise > > > and > > > >> you can miss peaks that fall outside of your sampling windows. But > > the > > > >> obvious benefit is that you don’t need a cluster to process these > > > streams. > > > >> You can probably perform most of your calculations with a > > multithreaded > > > >> java program. Sampling poses a few design challenges. First, where > > do > > > you > > > >> sample? Do you sample on the sensor? (the implication here is that > we > > > have > > > >> to program some sort of sampling capability in our sensors) . Do you > > > sample > > > >> on transport? (maybe a Flume interceptor or NiFi processor). There > is > > > also > > > >> a question of what the sampling rate should be. Not knowing > > statistical > > > >> properties of a stream ahead of time it’s hard to make that call. > > > >> > > > >> The third option I think is MR job. We can blast the data into HDFS > > and > > > >> then go over it with MR to derive the metrics we are looking for. > > Then > > > we > > > >> don’t have to sample or setup expensive infrastructure to receive a > > > deluge > > > >> of data. But then we run into the chicken and the egg problem that > in > > > >> order to size your HDFS you need to have data in HDFS. Ideally you > > > need to > > > >> capture at least one full weeks worth of logs because patterns > > > throughout > > > >> the day as well as every day of the week have different statistical > > > >> properties. So you need off peak, on peak, weekdays and weekends to > > > derive > > > >> these stats in batch. > > > >> > > > >> Any other design ideas? > > > >> > > > >> Thanks, > > > >> James > > > >> > > > >> > > > >> > > > >> > > > >> > > > >> On 4/13/16, 1:59 PM, "Nick Allen" <[email protected]> wrote: > > > >> > > > >> >If the tool starts at Kafka, the user would have to already have > > > committed > > > >> >to the investment in the infrastructure and time to setup the > sensors > > > that > > > >> >feed Kafka and Kafka itself. Maybe it would need to be further > > > upstream? > > > >> >On Apr 13, 2016 1:05 PM, "James Sirota" <[email protected]> > > > wrote: > > > >> > > > > >> >> Hi Goerge, > > > >> >> > > > >> >> This article defines micro-tuning of the existing cluster. What > I > > am > > > >> >> proposing is a level up from that. When you start with Metron > how > > do > > > >> you > > > >> >> even know how many nodes you need? And of these nodes how many > do > > > you > > > >> >> allocate to Storm, indexing, storage? How much storage do you > > need? > > > >> >> Tuning would be the next step in the process, but this tool would > > > answer > > > >> >> more fundamental questions about what a Metron deployment should > > look > > > >> like > > > >> >> given the number of telemetries and retention policies of the > > > >> enterprise. > > > >> >> > > > >> >> The best way to get this data (in my opinion) is to have some > tool > > > that > > > >> we > > > >> >> can plug into Metron’s point of ingest (kafka topics) and run > that > > > for > > > >> >> about a week or a month to be able to figure that out and spit > out > > > these > > > >> >> relevant metrics. Based on these metrics we can figure out the > > > >> fundamental > > > >> >> things about what metron should look like. Tuning would be the > > next > > > >> step. > > > >> >> > > > >> >> Thanks, > > > >> >> James > > > >> >> > > > >> >> > > > >> >> > > > >> >> > > > >> >> On 4/13/16, 9:52 AM, "George Vetticaden" < > > > [email protected]> > > > >> >> wrote: > > > >> >> > > > >> >> >I have used the following Kafka and Storm Best Practices guide > at > > > >> numerous > > > >> >> >customer implementations. > > > >> >> > > > > >> >> > > > >> > > > > > > https://community.hortonworks.com/articles/550/unofficial-storm-and-kafka-b > > > >> >> >est-practices-guide.html > > > >> >> > > > > >> >> > > > > >> >> >We need to have something similar and prescriptive for Metron > > based > > > on: > > > >> >> >1. What data sources are we enabling > > > >> >> >2. What enrichment services are we enabling > > > >> >> >3. What threat intel services are we enabling > > > >> >> >4. What are we indexing into Solr/Elastic and how long > > > >> >> >5. What are we persisting into HDFS.. > > > >> >> > > > > >> >> >Ideally, the The metron assessment tool combined with an > > > introspection > > > >> of > > > >> >> >the user’s ansible configuration should drive what ambari > > blueprint > > > >> type > > > >> >> >and configuration should be used when the cluster is spun up and > > the > > > >> storm > > > >> >> >topology is deployed. > > > >> >> > > > > >> >> > > > > >> >> >-- > > > >> >> >George VetticadenPrincipal, COE > > > >> >> >[email protected] > > > >> >> >(630) 909-9138 > > > >> >> > > > > >> >> > > > > >> >> > > > > >> >> > > > > >> >> > > > > >> >> >On 4/13/16, 11:40 AM, "George Vetticaden" < > > > [email protected] > > > >> > > > > >> >> >wrote: > > > >> >> > > > > >> >> >>+ 1 to James suggestion. > > > >> >> >>We also need to consider not just the data volume and storage > > > >> >> requirements > > > >> >> >>for proper cluster sizing but also processing requirements as > > well. > > > >> Given > > > >> >> >>that in the new architecture, we have moved to single > enrichment > > > >> topology > > > >> >> >>that will support all data sources, proper sizing of the > > enrichment > > > >> >> >>topology will be even more crucial to maintain SLAs and HA > > > >> requirements. > > > >> >> >>The following key questions will apply to each parser topology > > and > > > >> single > > > >> >> >>enrichment topology > > > >> >> >> > > > >> >> >>1. Number of workers? > > > >> >> >>2. Number of workers per machine? > > > >> >> >>3. Size of each workers (in memory)? > > > >> >> >>4. Supervisor memory settings > > > >> >> >> > > > >> >> >>The assessment tool should also be used to size topologies > > > correctly > > > >> as > > > >> >> >>well. > > > >> >> >> > > > >> >> >>Tuning Kafka, Hbase and Solr/Elastic should also be governed by > > the > > > >> >> Metron > > > >> >> >>assessment tool. > > > >> >> >> > > > >> >> >> > > > >> >> >>-- > > > >> >> >>George Vetticaden > > > >> >> >> > > > >> >> >> > > > >> >> >> > > > >> >> >> > > > >> >> >> > > > >> >> >> > > > >> >> >> > > > >> >> >>On 4/13/16, 11:28 AM, "James Sirota" <[email protected]> > > > wrote: > > > >> >> >> > > > >> >> >>>Prior to adoption of Metron each adopting entity needs to > > > guesstimate > > > >> >> >>>it¹s data volume and data storage requirements so they can > size > > > their > > > >> >> >>>cluster properly. I propose a creation of an assessment tool > > that > > > >> can > > > >> >> >>>plug in to a Kafka topic for a given telemetry and over time > > > produce > > > >> >> >>>statistics for ingest volumes and storage requirement. The > idea > > > is > > > >> that > > > >> >> >>>prior to adoption of Metron someone can set up all the feeds > and > > > >> kafka > > > >> >> >>>topics, but instead of deploying Metron right away they would > > > deploy > > > >> >> this > > > >> >> >>>tool. This tool would then produce statistics for data > > > >> ingest/storage > > > >> >> >>>requirement, and all relevant information needed for cluster > > > sizing. > > > >> >> >>> > > > >> >> >>>Some of the metrics that can be recorded are: > > > >> >> >>> > > > >> >> >>> * Number of system events per second (average, max, mean, > > > >> standard > > > >> >> >>>dev) > > > >> >> >>> * Message size (average, max, mean, standard dev) > > > >> >> >>> * Average number of peaks > > > >> >> >>> * Duration of peaks (average, max, mean, standard dev) > > > >> >> >>> > > > >> >> >>>If the parser for a telemetry exist the tool can produce > > > additional > > > >> >> >>>statistics > > > >> >> >>> > > > >> >> >>> * Number of keys/fields parsed (average, max, mean, > standard > > > dev) > > > >> >> >>> * Length of field parsed (average, max, mean, standard > dev) > > > >> >> >>> * Length of key parsed (average, max, mean, standard dev) > > > >> >> >>> > > > >> >> >>>The tool can run for a week or a month and produce these kinds > > of > > > >> >> >>>statistics. Then once the statistics are available we can > come > > up > > > >> with > > > >> >> a > > > >> >> >>>guidance documentation of recommended cluster setup. > Otherwise > > > it¹s > > > >> >> hard > > > >> >> >>>to properly size a cluster and setup streaming parallelism not > > > >> knowing > > > >> >> >>>these metrics. > > > >> >> >>> > > > >> >> >>> > > > >> >> >>>Thoughts/ideas? > > > >> >> >>> > > > >> >> >>>Thanks, > > > >> >> >>>James > > > >> >> >> > > > >> >> >> > > > >> >> > > > > >> >> > > > >> > > > > > > > > > > > > > > > >-- > > > >Nick Allen <[email protected]> > > > > > > > > > > > -- > > Nick Allen <[email protected]> > > > -- > > Jon > -- Nick Allen <[email protected]>
