On Thu, Feb 16, 2017 at 12:08 PM, Meier, Caleb <[email protected]> wrote: > Quick question. How timely is Fluo when it comes to processing > notifications? If there are enough workers, will a notification be processed > in a timely manner after writing to the observed column? Earlier we had a > discussion about a periodic query service. > If I write a notification to issue a periodic query to Fluo, can I expect > that my Observer will process that query fairly quickly (provided there are > enough workers)?
Fluo keeps track of the last time it saw a notification in a tablet and exponentially increases the scan period for that tablet when it keeps seeing nothing. The increase is up to a configurable maximum. The following code does the backoff. https://github.com/apache/incubator-fluo/blob/rel/fluo-1.0.0-incubating/modules/core/src/main/java/org/apache/fluo/core/worker/finder/hash/TabletData.java The following is where it gets the max sleep time. https://github.com/apache/incubator-fluo/blob/rel/fluo-1.0.0-incubating/modules/core/src/main/java/org/apache/fluo/core/worker/finder/hash/ScanTask.java#L72 Looking at the following, the default max sleep time for a tablet is 5 minutes. If I expanded the constant correctly, this can be changed by setting fluo.impl.ScanTask.maxSleep. Note, impl properties are not part of the public API and are subject to change when the implementation changes. https://github.com/apache/incubator-fluo/blob/rel/fluo-1.0.0-incubating/modules/core/src/main/java/org/apache/fluo/core/impl/FluoConfigurationImpl.java#L37 Also, when Fluo has notifications queued, it will wait till the queue size halves before scanning again for notifications. So if 10,000 notifications were queued, then scanning for notifications would not happen until the queue size was 5,000 or less. The following code shows where that happens. I noticed while looking for the max scan sleep code. https://github.com/apache/incubator-fluo/blob/rel/fluo-1.0.0-incubating/modules/core/src/main/java/org/apache/fluo/core/worker/finder/hash/ScanTask.java#L85 > > Caleb A. Meier, Ph.D. > Software Engineer II ♦ Analyst > Parsons Corporation > 1911 N. Fort Myer Drive, Suite 800 ♦ Arlington, VA 22209 > Office: (703)797-3066 > [email protected] ♦ www.parsons.com > > -----Original Message----- > From: Keith Turner [mailto:[email protected]] > Sent: Wednesday, February 01, 2017 11:03 PM > To: [email protected] > Subject: Re: third party service to poll Fluo for absence of event > > On Wed, Feb 1, 2017 at 9:54 PM, Christopher <[email protected]> wrote: >> On Wed, Feb 1, 2017 at 10:04 AM Meier, Caleb <[email protected]> >> wrote: >> >>> Yeah, this seems pretty reasonable to me. I guess it then boils down >>> to the nitty gritty of do I store results in Fluo and have my service >>> query Fluo (I think you guys actually advise against that in your >>> documentation), or export results and then have the service query >>> some external index that I am exporting to. >>> >>> >> I'm not sure we advise against it, so much as recognize that it may >> not be suitable for certain use cases and may not meet query >> performance expectations ( >> https://urldefense.proofpoint.com/v2/url?u=http-3A__fluo.apache.org_docs_fluo-2Drecipes_1.0.0-2Dincubating_export-2Dqueue_&d=CwIFaQ&c=Nwf-pp4xtYRe0sCRVM8_LWH54joYF7EKmrYIdfxIq10&r=vuVdzYC2kksVZR5STiFwDpzJ7CrMHCgeo_4WXTD0qo8&m=zqJSJTFo90FyUVCiF79uq3P0FHnxr0MLFKbsPsHGgyk&s=spmwJN_FBTO6TBBT2dne8sbE7MRMrlhz8lLPpfPZBbs&e= >> ). >> > > I would advise against querying Fluo for low latency queries. > However, this external service thats checking a few stats within Fluo and > injecting new notifications probably does not care about latency. > > The reason Fluo is not geared towards low latency is that it does lazy > recovery of failed transactions. Failed transactions are not cleaned > up until something tries to read the data, which could significantly delay > reads. > >> In any case, your observer need not write the final "last occurrence" >> entries into a Fluo table. It could write them anywhere. >> >> >>> Regarding timestamps, does the oracle server provide actual >>> timestamps or just logical timestamps? That is, could I use the >>> timestamps that the server provides to define some sort of now() >>> function to obtain the current time to compare with the times of incoming >>> events? >>> >> >> Just logical time, and it delivers batches to limit locking, so it can >> appear to jump ahead spontaneously. I'm not sure the OracleServer is >> suitable for this purpose. What level of precision are you going for? >> It might be enough to just run NTP, if you don't need more precision >> than "within seconds". >> >> >>> ________________________________________ >>> From: Christopher <[email protected]> >>> Sent: Tuesday, January 31, 2017 5:08 PM >>> To: [email protected] >>> Subject: Re: third party service to poll Fluo for absence of event >>> >>> You could write an observer which rolls up timestamps from all the >>> events you are concerned about, and puts the most recent event >>> timestamp into a centralized place, which you could poll. If there is >>> no ingest of these events, then the last timestamp in this central >>> place will exceed some threshold and the poller could detect that and >>> trigger additional actions. >>> >>> On Tue, Jan 31, 2017 at 3:51 PM Meier, Caleb >>> <[email protected]> >>> wrote: >>> >>> > Hello, >>> > >>> > I’m looking into using Fluo to develop an event based notification >>> > system that incrementally generates events of increasing >>> > complexity. The one issue that I’m running into is how to handle >>> > the non-event event. That >>> is, >>> > Fluo (as I understand it) is not well-suited to handle the >>> > following >>> > request: “generate a notification if no events of a given type have >>> > occurred within the last 24 hours”. This is because it is a push >>> > based notification framework that only generates notifications when >>> > things actually happen. So the question is, has anyone looked into >>> > developing a service for generating notifications at regular >>> > intervals (even if something doesn’t happen) that works with Fluo? >>> > I’m toying with the idea of creating some sort of Twill application >>> > that tells Fluo to wake up at regular intervals to generate a >>> > notification about the set of events falling within the given time >>> > window. Before doing this I just wanted to make sure that something >>> > like this does not already exist, and I also >>> want >>> > to get a sense of how bad an idea it is to delegate some of the logic of >>> > this periodic notification service to Fluo. Would it be better to >>> > separate out the temporal portion of my notification request to be >>> > processed entirely outside of Fluo to avoid transactional overhead? >>> > >>> > Caleb A. Meier, Ph.D. >>> > Software Engineer II ♦ Analyst >>> > Parsons Corporation >>> > 1911 N. Fort Myer Drive, Suite 800 ♦ Arlington, VA 22209 >>> > Office: (703)797-3066 <(703)%20797-3066> <(703)%20797-3066> >>> > [email protected]<mailto:[email protected]> ♦ >>> www.parsons.com< >>> > http://www.parsons.com/> >>> > >>> > -- >>> Christopher >>> >> -- >> Christopher
