Permalink
Switch branches/tags
Find file Copy path
362 lines (312 sloc) 20.6 KB

Prometheus Exposition Format: Protobuf vs. Text

by Björn “Beorn” Rabenstein, SoundCloud Ltd. @beorn7

Prometheus 1.x supports two exposition formats:

  • a Protocol Buffer format (“protobuf”).
  • a line-oriented text format (”text”).

Evolving the Prometheus exposition format(s) into a standard raises the question if both formats or only one of them should be part of the standard. Currently, the community gravitates towards the text format for a number of reasons. However, a number of aspects haven't been discussed widely so far. This document intends to raise awareness for them and turn the format decision into an informed one, whatever it will ultimately be.

Disclaimer: The author of this document joined the project in late 2013. While that's fairly early, he wasn't part of the project from the beginning. The claims about the “original vision” and similar as stated below are therefore not necessarily accurate but merely to the best of the author's knowledge. In the nascent project, a lot of knowledge was “tribal” and not yet documented in written form. Furthermore, the author was not involved in the recent development of Prometheus 2, which might lead to incorrect statements below. Those, at least, can be fixed easily. Corrections are welcome. On the other hand, the author was the lead designer of the text format. Thus, statements about original design decisions for the text format can be considered canonical.

Note about terminology: The terms histogram and summary in italics refer specifically to the Prometheus metric types. Their meaning has grown historically and is not necessarily in line with the general understanding of those terms. Counters and gauges, in contrast, are fairly well defined and no typographical distinction between the general terms and the Prometheus metric types is required.

Historical notes

Original role of the protobuf and text format

Note: Prometheus once also supported a JSON exposition format. The reasons for dropping it are discussed elsewhere. While interesting for the complete picture, they are not important for the discussion of protobuf vs. text format.

The protobuf format has a data model that is fundamentally richer than what Prometheus supports (even today with Prometheus 2.x):

  • It contains doc strings (help) per metric family.
  • It has first-class support for different metric types.
  • It has first-class support for complex metric types that cannot be represented with a single floating point number as a value:
    • Bucketed histograms.
    • Summaries with pre-calculated quantiles.

Prometheus throws away doc strings and the metric type and maps histograms and summaries into a number of simple gauge- or counter-like metrics (one per bucket of a histogram, one per quantile of a summary, one for the number of observations, one for the sum of observations).

The vision behind that was to create a rich “future-proof” data model, which Prometheus would later “grow into”. In that way, we could create a working prototype for Prometheus, covering the simple cases and harvesting the lowest hanging fruit, while instrumentation libraries can already support the rich data model, to the benefit of future Prometheus version or even non-Prometheus consumers.

The original motivation for the text format was twofold:

  • There are plenty of situations where the machinery of creating the protobuf wire format is not readily available, with the proverbial example being a shell script that needs to push a small number of samples to the Pushgateway, or a minimal Python program that intends to serve a small number of metrics via the HTTP server provided by the standard library. We call that use-case ad-hoc exposition or library-less exposition.
  • The data model of the protobuf format was often perceived as needlessly complicated, on the one hand because the Prometheus server was throwing away most of the hard work done before, on the other hand because a simple use-case only consisting of gauges and counters doesn't benefit much from the richness of the data model.

This resulted in two design principles for the text format that are very important to keep in mind:

  • The text format should be easy and fault-resistant to assemble, based on the assumption that it is, in most cases, either directly assembled by a human, or assembled programmatically but without much library support, i.e. conforming to strict format specifications would be a relevant burden. Minimizing efforts for the producer outweighs the efforts for the consumer.
  • The text format should work well for easy cases (a few gauges or counters) but doesn't have to deal well with complex cases (mostly histograms and summaries).

These design guidelines explain properties of the text format that are nowadays considered problematic:

  • The liberal use of whitespace everywhere.
  • The verbose and fragile representation of histograms and summaries. (In fact, the original code of the text format parser stated that it is in general a bad idea to represent histograms and summaries with the text format.)

Plenty of concerns were raised back then about the performance penalty of parsing text vs. decoding protobufs and the loss of robustness, consistency, and stringency. The way of breaking complex metric types into simple metric lines was the same as done by the Prometheus server to save the “rich” histograms and summaries in its “poor” embedded TSDB. This was explicitly meant as convenient coincidence and not as a leak of Prometheus implementation details into the exposition format. To make sure of that, the text format parser was implemented to convert the text format into protobuf (rather than the internal data model of the Prometheus server), so that we could still follow our vision of growing the internal data representation into something closer to the rich data model of the protobuf format.

The text format was also designed to closely match the appearence of metric selectors in PromQL.

The design efforts paid off soon as the text format became popular quite quickly, with some intended and some not so intended side effects:

  • The text format was used as the usual human-readable representation of metrics when looking at the /metrics endpoint with a browser. (An easy to implement alternative would have been to show the text representation of the protobuf format. But that would have been quite hard to read for humans, especially in simple cases.)
  • It enabled the textfile collector in the Node Exporter, one of the most convenient ways of exposing metrics: Drop a snippet with metrics in the text format into a designated directory, and the Node Exporter will immediately start to expose it. Many questions of “How to export X?” could now be answered by one line of Bash run by Cron.
  • The maintainer of the Python and Java client libraries came to the conclusion that a dependency on protobuf causes issues in those languages and would impede adoption. Thus, he happily implemented those libraries without protobuf support. (In case of the Java library, his rewrite had no protobuf support while the original, quite prototypical implementation still had it.) Other libraries mostly followed that example, with the notable exception of the Go client library (which is also the oldest library). The Go library follows the spirit of the original vision of approaching the protobuf data model to the fullest possible extent: It internally directly acts on protobuf messages. Only if explicitly asked for the text format (e.g. via an HTTP request by a browser), it converts the protobuf representation into the text format.

The performance overhead of the text format was in general considered as small enough to not worry about client libraries without protobuf support.

Recent developments

For the upcoming Prometheus 2, the internal storage was completely rewritten, with dramatic performance improvements along almost all dimensions. The rewrite went along with a new internal data model for metrics. However, this didn't follow the original vision of approaching the rich protobuf data model. On the contrary, it was in embracing the text format instead, with various beneficial results.

An important incentive was performance analysis. As it turned out, the whole decoding chain from the protobuf wire-format into the internal data model of Prometheus was quite wasteful. In Prometheus 1.x, starting the chain at the text format was unsurprisingly even more wasteful because the text parser was parsing the text format into protobufs first, as described above.

For Prometheus 2, an allocation-free text-format decoder was written that decodes the text format directly into the (new and improved) internal data model. As it turned out, this decoding chain is dramatically more efficient than the original protobuf chain. Not without historical irony, things changed completely, from “You can use the text format, but only the protobuf format will give you really high performance.” to “You should avoid the protobuf format, it eats quite a few of your cores.”

On top of that, Prometheus 2 avoids calculating a metric hash on each ingestion. This tweak also leverages how the text format is laid out on the wire. In high-ingestion scenarios, metrics hashing would otherwise have become the dominant CPU consumer.

Comparison of the protobuf and text format

This section compares different aspects and implications of both formats.

General applicability and upgrade paths

It is important to keep in mind that protobuf encoding and decoding is not necessarily fast (as was naively assumed at the time the text format was created and concerns were raised that the text format decoding would be too slow). A hand-written encoder/decoder for a custom format might easily outperform the code auto-generated by the protobuf compiler. (But see the section about performance below for further considerations.) Arguably, the key advantage of protocol buffers is to provide the following featurs at a decent encoding/decoding performance:

  1. Create encoders and decoders in all supported languages from a single .proto file.
  2. Changes of the format are therefore easy to push (the .proto file is the only code to change manually). Plus, protocol buffers are designed to allow forward and backward compatibility.
  3. The structured data as described by the .proto file can be directly operated on in the target language. There is no need to translate to and from the internal data model. The protobuf description is the data model.

One might argue that each of the items above are of limited relevance for Prometheus:

  1. Encoders for both the protobuf and text format exist in various languages already (in the form of Prometheus client libraries). For some languages, there are even decoders available. (However, the choice of languages for which protobuf compilers exist is still richer than the choice of languages for which text format encoders exist.)
  2. With the maturity of the Prometheus ecosystem, future changes of the exposition format are expected to be rare. (This expectation might be seen as naive, though.)
  3. Prometheus so far has failed to benefit from this item, as the internal data model is different from the protobuf description in both Prometheus 1.x and 2.x (which is arguably either a shortcoming of Prometheus or of the protobuf representation in Go, see also the section about the data model below).

The arguments above are based on a “Prometheus only” scenario. Obviously, the whole point of evolving the exposition format into a standard is applicability outside of the Prometheus ecosystem, too. Viewing each item from that perspective changes the outcome considerably:

  1. The ability to easily create encoders and in particular decoders in many languages, as provided by the protobuf format, becomes more important.
  2. As the standardization will add more requirements for the exposition format (e.g. timestamps of sub-millisecond precision, different value types, …), more changes are expected in the process.
  3. Encoders and decoders created for Prometheus encode from and decode to the internal Prometheus data model. This might not be the data model non-Prometheus users want to use. See the section about the data model below for details.

In summary, the advantages of a protobuf format are more relevant for a widely adopted standard than for a format meant to be used predominantly within the Prometheus ecosystem.

Implied data model

As described above, the internal Prometheus data model is different from the data model of the protobuf format. The original reason was that Prometheus had to reach the state of a working prototype, and more complex metric representation was postponed. However, there is also a problem with the code the official Go protobuf compiler generates. It's just not very idiomatic for Go, so acting on the generated data structures as the internal data model is cumbersome (as demonstrated in the Prometheus Go client library). This has somewhat improved with proto3, but Prometheus hasn't moved to proto3 yet, see this GH issue for details. Another option would be an alternative Go protobuf implementation like gogo/protobuf, which arguably generates more idiomatic Go code and would make it easier to act directly on the protobuf data structures.

For the reasons stated above, the text format follows the internal Prometheus data model much more closely than the protobuf format. While explicitly not intended as such, the text format effectively leaks the internal Prometheus model into the exposition format. The main implication is that complex metric types (histogram, summary) are broken down into individual metric lines with 64bit floating point numbers as data type and some “magic” labels (le, quantile). This is verbose and prone to inconsistencies. For example, the text format formally allows to specify a timestamp per line. In a histogram and summary, however, all quantiles, buckets, and the sum and count of observations must have the same timestamp. Furthermore, data types are not appropriate anymore. The count of a bucket and the count of observations are supposed to be unsigned integers but are represented as floating point numbers in the text format. The value of the le label is supposed to be a floating point number but is represented as an UTF-8 string. The protobuf format does not have any of these issues.

A different aspect is that the representation of histograms as naive buckets is most certainly not the end of the journey. Many ideas about dynamic bucketing and/or inexpensive representation of higher resolution are floating around, which would result in representing some kind of digest or even binary compressed data (more research needed, but this is definitely a very hot topic with huge potential for future uses). The text format in its current form needed to change fundamentally to accommodate such innovations.

Highly relevant here are the comments by Sumeer Bhola and Jeromy Carriere. An important conclusion is that flattening of a complex metric type is not the responsibility of the exposition format but that of the consuming backend.

It should also be noted here that the current way how Prometheus handles complex metric types has a lot of issues even with a completely Prometheus-centric view. See this slide deck from the Prometheus developer summit Berlin 2017 for details. Since a change within Prometheus is highly desirable, we could, ironically, end up in a state where the standardized text format would still follow the leaked internal data model, even after the internal data model of Prometheus has improved.

There is the option of changing the text format to not break down complex metric type. A histogram sample could look like this:

some_histogram{foo="bar"} {sum: 47.11, count: 42, 0.1: 1, 0.2: 5, 0.5: 12, 1: 37}

This would solve many of the issues above, but would still make it quite hard to change to advanced representations of histograms. Lines could also become very long with many buckets. See also the section about performance.

Verbosity and compression

The text format is needlessly verbose for histograms and summaries. The impact is particularly high in a use-case with many buckets on a highly dimensional metric. Since the data is mostly redundant, it compresses nicely. In the Prometheus context, scrapes are gzip-compressed by default, which results in no noticable size difference between text and protobuf scrapes. However, there are scenarios where the resources needed for compression might be a relevant burden. The most likely case is a resource-tight monitoring target. But with the dramatic improvements in ingestion performance featured by Prometheus 2.x, compression might become an issue on the side of the Prometheus server, too (or other consumers, as we are talking about making the exposition format generally useful here, not just for Prometheus).

A change of the text format as suggested in the previous section, where buckets of a histogram and quantiles of a summary are all listed on the same line, would help here. But see also the next section about performance.

Ingestion performance and the benefits of a binary encoding

As reported above, the new hand-coded text format decoder in Prometheus 2.x is dramatically more efficient than the protobuf decoding in Prometheus 1.x. Also, the text format was leveraged for ingestion tweaks, saving time on hashing metrics and labels. This result deserves a closer look.

For one, there is the aspect of a hand-coded vs. a generated decoder. If maximum performance is the first priority, a protobuf decoder could be hand-coded, too. The necessity of hand-coding a decoder for the text format should not be taken as an advantage for the text format. While hand-coding a protobuf decoder would be quite extreme, the ingestion tweaks implemented on top of the on-the-wire representation of the text format could fundamentally implemented in the same way on top of the protobuf wire format.

Then there are the known problem with the official Go protobuf compiler. Not only is the generated code not very idiomatic, as discussed above, it is also quite inefficient. Alternative protobuf implementation perform way better, notably gogo/protobuf. See this CloudNativeCon talk for a case study of protobuf decoding performance of the official protobuf compiler vs gogo/protobuf.

Finally, the current payload is mostly strings, which are naturally encoded in more or less the same way in a text-based format vs. protobuf. The encoding of numbers is fundamentally less efficient in a text-based format. One of the pain points with Prometheus right now is the relatively high cost of a histogram bucket. Users have to be very judicious with their bucketing schemes. With an improved Prometheus (or with another backend that deals differently with histograms), we will see payloads dominated by histogram buckets – or by more efficient representations thereof, i.e. some kind of digest or binary compressed data. In either case, it will tip the balance from string-dominated towards mostly numeric or even binary data, for which a text-based format is not a natural fit.

Required changes for standardization

The above suggests that the move from a Prometheus-targeted format to a generally useful metrics exposition format will require many changes of the text format. However, it should be noted that also the protobuf format will require some changes, too. This document is not meant to imply that all concerns will go away by choosing the protobuf format. However, there are fewer fundamental problems, and the protobuf format is better suited for ongoing improvements and additions of the format.

Implications for gRPC

With the rise of gRPC, the request for exposing metrics via gRPC was inevitable. While gRPC is in principle payload-agnostic, it integrates most naturally with protocol buffers. We would essentially send structured data as an opaque strings via gRPC if we used the text format.