Re: [prometheus-developers] Re: Remote Write Metadata propagation
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 whate
Re: [prometheus-developers] Re: Remote Write Metadata propagation
Thanks Rob for putting this proposal together, I think it highlights some features of what we want metadata RW and remote write in general to look like in the future. As others have pointed out (thanks Björn for giving such a detailed description) there's issues with the way Prometheus currently handles metadata that need to be thought about and handled differently when storing metadata in the WAL or in long term storage. I didn't make many more comments as most of what I wanted to say had already been mentioned by others. As part of thinking about how to get metadata and exemplars into remote write, some of us have been discussing what we've been calling 'the future of remote write'. While there's nothing formal yet, I will be starting a brainstorming/design doc soon and would appreciate your input there Rob. -- 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 prometheus-developers+unsubscr...@googlegroups.com. To view this discussion on the web visit https://groups.google.com/d/msgid/prometheus-developers/CAN2d5OTRMVzAqZ%3DLqpSv6bx9Z%2Btvw3MZCmBxj4e8B1T1YOJr6Q%40mail.gmail.com.
Re: [prometheus-developers] Re: Remote Write Metadata propagation
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 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 i
Re: [prometheus-developers] Re: Remote Write Metadata propagation
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 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 > 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