Hey Callum, Apologies missed your response as was typing back to Björn.
Look forward to seeing your document, sounds good. As I mentioned in my previous email I think that there's definitely "further work" area. I'd like to get just TYPE and if it's not too difficult then HELP and UNIT at least too flowing sooner than that timeline however, and we have folks ready to contribute to work in this space right now. Would love to hear your thoughts on my latest proposal as sent with the last email. Best, Rob On Thu, Aug 6, 2020 at 5:58 PM Rob Skillington <[email protected]> wrote: > Hey Björn, > > > Thanks for the detailed response. I've had a few back and forths on this > with > Brian and Chris over IRC and CNCF Slack now too. > > I agree that fundamentally it seems naive to idealistically model this > around > per metric name. It needs to be per series given what may happen w.r.t. > collision across targets, etc. > > Perhaps we can separate these discussions apart into two considerations: > > 1) Modeling of the data such that it is kept around for transmission > (primarily > we're focused on WAL here). > > 2) Transmission (and of which you allude to has many areas for > improvement). > > For (1) - it seems like this needs to be done per time series, thankfully > we > actually already have modeled this to be stored per series data just once > in a > single WAL file. I will write up my proposal here, but it will surmount to > essentially encoding the HELP, UNIT and TYPE to the WAL per series similar > to > how labels for a series are encoded once per series in the WAL. Since this > optimization is in place, there's already a huge dampening effect on how > expensive it is to write out data about a series (e.g. labels). We can > always > go and collect a sample WAL file and measure how much extra size > with/without > HELP, UNIT and TYPE this would add, but it seems like it won't > fundamentally > change the order of magnitude in terms of "information about a timeseries > storage size" vs "datapoints about a timeseries storage size". One extra > change > would be re-encoding the series into the WAL if the HELP changed for that > series, just so that when HELP does change it can be up to date from the > view > of whoever is reading the WAL (i.e. the Remote Write loop). Since this > entry > needs to be loaded into memory for Remote Write today anyway, with string > interning as suggested by Chris, it won't change the memory profile > algorithmically of a Prometheus with Remote Write enabled. There will be > some > overhead that at most would likely be similar to the label data, but we > aren't > altering data structures (so won't change big-O magnitude of memory being > used), > we're adding fields to existing data structures that exist and string > interning > should actually make it much less onerous since there is a large > duplicative > effect with HELP among time series. > > For (2) - now we have basically TYPE, HELP and UNIT all available for > transmission if we wanted to send it with every single datapoint. While I > think > we should definitely examine HPACK like compression features as you > mentioned > Björn, I think we should think more about separating that kind of work > into a > Milestone 2 where this is considered. For the time being it's very > plausible > we could do some negotiation of the receiving Remote Write endpoint by > sending > a "GET" to the remote write endpoint and seeing if it responds with a > "capabilities + preferences" response, and if the endpoint specifies that > it > would like to receive metadata all the time on every single request and > let > Snappy take care of keeping size not ballooning too much, or if it would > like > TYPE on every single datapoint, and HELP and UNIT every DESIRED_SECONDS or > so. > To enable a "send HELP every 10 minutes" feature we would have to add to > the > datastructure that holds the LABELS, TYPE, HELP and UNIT for each series a > "last sent" timestamp to know when to resend to that backend, but that > seems > entirely plausible and would not use more than 4 extra bytes. > > These thoughts are based on the discussion I've had and the thoughts on > this > thread. What's the feedback on this before I go ahead and re-iterate the > design > to more closely map to what I'm suggesting here? > > Best, > Rob > > On Thu, Aug 6, 2020 at 2:01 PM Bjoern Rabenstein <[email protected]> > wrote: > >> On 03.08.20 03:04, Rob Skillington wrote: >> > Ok - I have a proposal which could be broken up into two pieces, first >> > delivering TYPE per datapoint, the second consistently and reliably >> HELP and >> > UNIT once per unique metric name: >> > >> https://docs.google.com/document/d/1LY8Im8UyIBn8e3LJ2jB-MoajXkfAqW2eKzY735aYxqo >> > /edit#heading=h.bik9uwphqy3g >> >> Thanks for the doc. I have commented on it, but while doing so, I felt >> the urge to comment more generally, which would not fit well into the >> margin of a Google doc. My thoughts are also a bit out of scope of >> Rob's design doc and more about the general topic of remote write and >> the equally general topic of metadata (about which we have an ongoing >> discussion among the Prometheus developers). >> >> Disclaimer: I don't know the remote-write protocol very well. My hope >> here is that my somewhat distant perspective is of some value as it >> allows to take a step back. However, I might just miss crucial details >> that completely invalidate my thoughts. We'll see... >> >> I do care a lot about metadata, though. (And ironically, the reason >> why I declared remote write "somebody else's problem" is that I've >> always disliked how it fundamentally ignores metadata.) >> >> Rob's document embraces the fact that metadata can change over time, >> but it assumes that at any given time, there is only one set of >> metadata per unique metric name. It takes into account that there can >> be drift, but it considers them an irregularity that will only happen >> occasionally and iron out over time. >> >> In practice, however, metadata can be legitimately and deliberately >> different for different time series of the same name. Instrumentation >> libraries and even the exposition format inherently require one set of >> metadata per metric name, but this is all only enforced (and meant to >> be enforced) _per target_. Once the samples are ingested (or even sent >> onwards via remote write), they have no notion of what target they >> came from. Furthermore, samples created by rule evaluation don't have >> an originating target in the first place. (Which raises the question >> of metadata for recording rules, which is another can of worms I'd >> like to open eventually...) >> >> (There is also the technical difficulty that the WAL has no notion of >> bundling or referencing all the series with the same metric name. That >> was commented about in the doc but is not my focus here.) >> >> Rob's doc sees TYPE as special because it is so cheap to just add to >> every data point. That's correct, but it's giving me an itch: Should >> we really create different ways of handling metadata, depending on its >> expected size? >> >> Compare this with labels. There is no upper limit to their number or >> size. Still, we have no plan of treating "large" labels differently >> from "short" labels. >> >> On top of that, we have by now gained the insight that metadata is >> changing over time and essentially has to be tracked per series. >> >> Or in other words: From a pure storage perspective, metadata behaves >> exactly the same as labels! (There are certainly huge differences >> semantically, but those only manifest themselves on the query level, >> i.e. how you treat it in PromQL etc.) >> >> (This is not exactly a new insight. This is more or less what I said >> during the 2016 dev summit, when we first discussed remote write. But >> I don't want to dwell on "told you so" moments... :o) >> >> There is a good reason why we don't just add metadata as "pseudo >> labels": As discussed a lot in the various design docs including Rob's >> one, it would blow up the data size significantly because HELP strings >> tend to be relatively long. >> >> And that's the point where I would like to take a step back: We are >> discussing to essentially treat something that is structurally the >> same thing in three different ways: Way 1 for labels as we know >> them. Way 2 for "small" metadata. Way 3 for "big" metadata. >> >> However, while labels tend to be shorter than HELP strings, there is >> the occasional use case with long or many labels. (Infamously, at >> SoundCloud, a binary accidentally put a whole HTML page into a >> label. That wasn't a use case, it was a bug, but the Prometheus server >> ingesting that was just chugging along as if nothing special had >> happened. It looked weird in the expression browser, though...) I'm >> sure any vendor offering Prometheus remote storage as a service will >> have a customer or two that use excessively long label names. If we >> have to deal with that, why not bite the bullet and treat metadata in >> the same way as labels in general? Or to phrase it in another way: Any >> solution for "big" metadata could be used for labels, too, to >> alleviate the pain with excessively long label names. >> >> Or most succintly: A robust and really good solution for >> "big" metadata in remote write will make remote write much more >> efficient if applied to labels, too. >> >> Imagine an NALSD tech interview question that boils down to "design >> Prometheus remote write". I bet that most of the better candidates >> will recognize that most of the payload will consist of series >> indentifiers (call them labels or whatever) and they will suggest to >> first transmit some kind of index and from then on only transmit short >> series IDs. The best candidates will then find out about all the >> problems with that: How to keep the protocol stateless, how to re-sync >> the index, how to update it if new series arrive etc. Those are >> certainly all good reasons why remote write as we know it does not >> transfer an index of series IDs. >> >> However, my point here is that we are now discussing exactly those >> problems when we talk about metadata transmission. Let's solve those >> problems and apply them to remote write in general! >> >> Some thoughts about that: >> >> Current remote write essentially transfers all labels for _every_ >> sample. This works reasonably well. Even if metadata blows up the data >> size by 5x or 10x, transfering the whole index of metadata and labels >> should remain feasible as long as we do it less frequently than once >> every 10 samples. It's something that could be done each time a >> remote-write receiver connects. From then on, we "only" have to track >> when new series (or series with new metadata) show up and transfer >> those. (I know it's not trivial, but we are already discussing >> possible solutions in the various design docs.) Whenever a >> remote-write receiver gets out of sync for some reason, it can simply >> cut the connection and start with a complete re-sync again. As long as >> that doesn't happen more often than once every 10 samples, we still >> have a net gain. Combining this with sharding is another challenge, >> but it doesn't appear unsolveable. >> >> -- >> Björn Rabenstein >> [PGP-ID] 0x851C3DA17D748D03 >> [email] [email protected] >> > -- You received this message because you are subscribed to the Google Groups "Prometheus Developers" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. To view this discussion on the web visit https://groups.google.com/d/msgid/prometheus-developers/CABakzZa7SjVY02kkHm4N6kidC-WrL8HBoKvtLJaWAJOgGMSLOQ%40mail.gmail.com.
