This quarter I worked on creating a command line interface for signature generation and in doing that extracted it from the processor into a standalone-ish module.
The end result of this work is that:
- anyone making changes to signature generation can can test the changes out on their local machine using a Socorro local development environment
- I can trivially test incoming signature generation changes--this both saves me time and gives me a much higher confidence of correctness without having to merge the code and test it in our -stage environment 
- we can research and experiment with changes to the signature generation algorithm and how that affects existing crash signatures
- it's a step closer to being usable by other groups
This blog post talks about that work briefly and then talks about some of the things I've been able to do with it.
|||I can't overstate how awesome this is.|
What is Socorro?
First off, what is Socorro?
Socorro is the crash ingestion pipeline for Firefox and other Mozilla products.
When Firefox crashes, the crash reporter asks the user if the user would like to send a crash report. If the user answers "yes!", then the crash reporter collects data related to the crash, generates a crash report, and submits that crash report as an HTTP POST to Socorro. Socorro saves the crash report, processes it, and has tools for viewing and analyzing crash data.
What is signature generation?
For every crash report that is processed, Socorro creates a signature for that crash report. The algorithm for generating a signature consists of an intricate series of gestures and transformations honed over many years codified as a pipeline of rules and several lists.
For example, a crash report for an out-of-memory crash might end up with a signature like this:
OOM | large | mozalloc_abort | mozalloc_handle_oom | moz_xmalloc | std::_Allocate | std::vector<T>::_Reallocate
Another related to GPU might look like this:
We use signatures for a couple of things: categorization and correlation.
We use signatures to categorize crashes into buckets in order to discern signal from the noise. We get a lot of crashes and we can't look at all of them. What we really want to do is look at important ones:
- security issues
- crashes that heavily affect a lot of people
- crashes that heavily affect a specific group of people (people using a right-to-left language, people using a specific GPU, people using a specific version of Windows, people using a specific antivirus plugin, and so on)
- crashes on a specific channel (release, beta, nightly)
As such, Socorro's analysis tools look at groups of crashes rather than individual crashes. Signatures let us group crashes by rough indication of the cause of the crash.
For example, the Top Crashers report shows top crash signatures for some period and movement in rank since the last period. This report helps us prioritize engineering work and in the case of the nightly and beta channels, helps us suss out bugs in recent changes.
We use signatures to correlate groups of crashes between different data sets. This should theoretically let us analyze data trends in one system, then use signatures to look up group of crashes in other systems. This would be super helpful because Mozilla has different sets of crash data that come from different places and connecting these data sets will surface issues we can't otherwise see.
Over the last year, I've had several conversations with people who had data sets they wanted to correlate with Socorro's crash data.
Late last year, David and team were trying to do client-side symbolification and signature generation for crash ping data in Telemetry. Signature generation would allow them to analyze change in rates on a different data set that had different properties than Socorro's, but then take the fruits of that analysis and examine the bigger issues in a more detailed way with Socorro data.
Adam spent some time copying Socorro signature generation code and distilling it down so that it could run in an IPython notebook. That was probably a ton of work and great for a prototype, but completely not viable for long-term usage.
Christian built a separate crash ingestion system for collecting crashes from fuzzing Firefox. These crashes don't come from real users, so it's good to have them in a separate system. He really wanted to correlate their data with Socorro data to see if crashes they were seeing from fuzzing existed in the wild. These could be possible security problems.
Socorro signatures are unique to our system plus they're constantly changing. Anyone who wanted to correlate using signatures would need to use our code and since our code is impossible to use or extract, that makes this a very sad story currently.
Current state of Socorro's signature generation
Socorro is like 10 years old. Over that period of time, we've honed signature generation to fix signatures that were too big and too small and handle changes in Firefox architecture, libraries, compilers, platforms, and additional products.
The signature generation system is deeply embedded into the Socorro processor. It consists of a set of processor transform rules that take in raw and processed crash data and adjust the signature accordingly. The rules look at many parts of the crash report and the processed crash report to create a signature: the stack of the crashing thread, classifications, IPC channel name and error (if there was one), status strings, and other things. There's no way to run signature generation on a crash without also processing it and because it's intricately tied into the processor, it's impossible to use outside of Socorro.
There's a spectrum of signature quality: good to bad.
A good signature is one that puts crashes that have differing symptoms but the same cause into a single group with a single signature.
A bad signature is one that puts crashes that don't really have anything to do with one another in the same group. Maybe it groups a bunch of crashes that happened on a Thursday in June. Nice, but unlikely to be helpful. Bad signatures can also create groups that are "too big" (too many unrelated crashes) or groups that are "too small" (many signatures cover the same problem).
Signatures are in constant flux and can be affected by many things: new versions of operating systems system libraries, changes to libraries Firefox uses, and changes to Firefox (contents of crash report, JIT changes, architecture changes, and so on).
Today's crashes will probably look different than crashes from 6 months ago even if they're for the same problem. However, the better our signatures are, the more likely we can see causes of crashes over long periods of time.
We're constantly tweaking signature generation. Most changes happen in files that contain lists of regular expressions for tokens that show up in stack frames. We call these "siglists" for reasons I'm not entirely sure of. Maybe it's short for "signature lists" or something. We have a list for tokens we never care about. We have a list for tokens that should be part of the signature and indicate we should continue to the next frame of the stack. I think we average a change a week, but they tend to come in spurts.
When someone goes to make a change to one of the lists, they open up a pull request with the change. There's no way to test the change before opening a pull request. There's no way to test the change during review, either. Generally, we eyeball the change, harumph a few times, merge it, and then test it once it's been deployed to our -stage environment by reprocessing a few selected crashes.
In summary, our signature generation system has been honed over many years, is intricately tied to the processor, consists of many intricate steps, changes often, and doesn't have a good workflow for changing it.
Why work on it now?
Adrian used to review all incoming signature generation pull requests, but then he went off to a different team and I took over his duties. He was familiar with signature generation. I was not. So there I was trying to review changes with no experience/wisdom with how changes can affect signatures.
Lonnen wanted me to look at other ways we could generate signatures particularly whether generating signatures using less of the raw and processed crash was viable. In order to do that, I'd have to build a harness of some kind that didn't exist, yet.
Lastly, I had finished up extracting the Socorro collector (now called Antenna) into a separate project and was looking at how to do the same thing for the processor. The processor is more complicated, so I was thinking it might be nice to extract signature generation first and thus break it up into a series of smaller, more impactful steps.
Relatedly, I had conversations with other groups about correlating using signatures and it's a sad story. There's a lot of work that we'd need to do to fix that. Seems like a good idea to take a step in that direction now while I'm making it easier to test changes and experiment.
All of that led up to me throwing together a prototype to prove viability of this project and then spending a quarter on extracting signature generation and throwing a command line interface on it.
I set out to make the following things possible:
People making changes can test changes in a local development environment. People can verify their changes are correct before making a pull request. I can verify correctness when merging. I can also test changes on 1000s of crashes and see if they affect things in adverse ways.
Fixing this will improve our ability to make signature generation changes quickly and accurately and improve our confidence for correctness. Because signatures are one of the cornerstones for most of Socorro's analysis tools, this is a big deal.
I want to experiment with signature generation changes. I want to be able to test changes on 1000s of crashes and understand what changed. This makes it possible to do projects like "make signatures for this group of crashes better". Right now that kind of project is like "shovel the driveway with this spoon".
I want to be able to test changes on crashes that meet certain criteria. For example, testing signature changes on all the crashes for Firefox 58 in the nightly channel that are shutdownhangs.
We need tools to be able to experiment with the signature generation system in order to improve it.
I want to be able to support multiple signatures. During processing, Socorro generates two signatures: a "proto signature" which is based entirely on the stack frames of the crashing thread and a "signature" which is generated from the signature generation algorithm. Both of these are created with the current signature generation system.
Signature generation is a static pipeline of rules. Socorro can only create these two signatures and no others during processing.
I want to turn signature generation into its own pipeline so that we can run multiple pipelines and get multiple signatures. This opens up some possibilities for correlation in the future.
We want to make it easier for other groups to use our signature generation code. Socorro's signatures are kind of like Nonna's special sauce--you can't make them anywhere else. It's impossible to correlate data between systems if there's no common key. If we want to use the Socorro signature as the common key, then other groups need to be able to generate that from their crash data.
Reduce maintenance burden. Socorro is huge. We need to make changes that enable us to do more, better, with less resources. This change has to reduce maintenance burden.
What I did
In July, I wrote a prototype. I wanted to know if I could build a signature generation system that could generate signatures for crashes using data that was publicly available from the Socorro APIs. The prototype took me a couple of days to build. That let me show Lonnen that the idea had legs, plus I could show him all the fancy things I could do with it even though it was a rough prototype.
The prototype proved the project was viable, so I wrote up bug #1385970 which focused on writing a signature generation command line interface with extracting signature generation being a side-effect of working on the bug.
I was concerned about how big the project was, so I spent time reducing the scope. Having a standalone library would be great, but that's a lot of additional up-front work and ongoing maintenance work. I pushed standalone library work off to a later phase, but kept this in mind as I worked on the project.
We had a bunch of other big projects in the queue for Socorro, so this project simmered on the back burner and when I had time to work on it, I ended up rewriting it a couple of times. The prototype was rough, so this was for the best.
I spent a lot of time landing smaller changes that could stand on their own. This reduced the size of the final pull request making it smaller, easier to review, and less risk. It also meant I was making progress by shipping small changes. However, it probably added to the overall length of the project.
I cleaned up the tests for signature generation rules.
I converted the processor transform rules into signature transform rules so generating signatures no longer depends on the rest of the processor.
I simplified the rule structure, too, since we didn't need to support requirements the rest of the processor had. I wish I had simplified it more--it feels clunky now.
I removed use of DotDict and configman and other Socorro-y things from signature generation code. It still uses some things from Socorro, but only minor ones. I tried to minimize the number of libraries required. This will reduce the work required to fully extract the library when we eventually do that.
I moved the SignatureJitCategory rule to be with the rest of the signature generation rules. That rule originally ran later in the processor pipeline, but if we wanted to generate a signature as a single step, I needed to move that earlier to be with the other signature generation rules.
I wrapped the whole thing in a command line interface that can take crash ids from stdin or the command line. (More on that in a bit.)
I went through and made it a bit more Python 3 friendly. We don't run it in a Python 3 environment and there's no Python 3 testing, but it's more likely to work in Python 3 now.
I wrote a couple of "outputters" for the command line interface so it can spit out output in text as well as CSV. I figured those were the formats that help me do my work. We can easily add others if someone needs them.
The new command line interface
Let's talk about what we have now.
The command line interface runs in the processor container of the Socorro local development environment. You could build a different container to run it in--it just needs certain Python libraries to be installed.
It has a usage like this:
python -m socorro.signature CRASHID [CRASHID ...]
You can get help like this:
(container) $ python -m socorro.signature --help
It can format the output in text:
(container) $ python -m socorro.signature eabeedd9-5e6f-4e03-a7f2-220a70171004 Crash id: eabeedd9-5e6f-4e03-a7f2-220a70171004 Original: mozalloc_abort | abort | core::option::expect_failed | webrender::frame_builder::FrameBuilder::build New: mozalloc_abort | abort | core::option::expect_failed | webrender::frame_builder::FrameBuilder::build Same?: True
Or CSV (long line that I wrapped):
python -m socorro.signature --format=csv eabeedd9-5e6f-4e03-a7f2-220a70171004 "crashid","old","new","same?","notes" "eabeedd9-5e6f-4e03-a7f2-220a70171004","mozalloc_abort | abort | core::option::expect_failed | webrender::fram e_builder::FrameBuilder::build","mozalloc_abort | abort | core::option::expect_failed | webrender::frame_build er::FrameBuilder::build","True",""
When you're working on changes and want to see how changes affect a lot of crashes, you can use the --different-only flag to show just the signatures that changed.
It can take crash ids from stdin as well as from the command line. You can use it in concert with scripts/fetch_crashids.py to do a Super Search in Crash Stats for signatures that contain some string and run all those through signature generation. For example, this looks at all signatures for the Firefox product from yesterday that have "shutdownhang" in the signature:
(container) $ ./scripts/fetch_crashids.py --signature-contains=shutdownhang | \ python -m socorro.signature
Since I landed this, I've used it for reviewing pull requests. Reviews used to be a conversation like this one:
We think it'll change it in the way we want, but have no idea until we've merged and it's deployed to -stage. Fixing it would involve another PR plus we'd have to either revert the change or fix it before deploying to -prod.
Fully reviewing this (including testing on -stage) would take an hour or more depending on how long it took to deploy to -stage.
Now it's a conversation like this one:
Reviewing takes several minutes and we no longer merge anything that blocks a -prod deploy. We can also look at changes for lots of crashes if the situation warrants it.
I also worked on bug #1402037 which aimed to improve signatures for shutdownhang crashes which exceeded 255 characters and thus were truncated. One of the things I did for that bug was make a big text file with all the crash ids for shutdownhang crashes over the last week. There were around 17,000 crash ids in that list.
I built the list like this:
(container) $ ./scripts/fetch_crashids.py --url=REALLYLONGSUPERSEARCHURL > bug1402037.txt
Then I ran signature generation like this:
(container) $ cat bug1402037.txt | python -m socorro.signature --format=csv > output.csv
Then I'd pull the file into LibreOffice Calc and sort it and see what happened.
The signature generation command line interface makes that easy. However, it takes a really long time to run because for each crash it's pulling raw and processed information from the Crash Stats APIs. I wrote up bug 1403339 for caching raw and processed data which would radically speed up running signature generation a second time on the same set of crash ids.
There's more information about signature generation in the signature generation documentation.
Prototypes grease the wheels of progress. The prototype proved the project was viable and worth working on.
It's not quite perfect especially regarding JIT-related crashes. There's a JIT classification annotation in the processed crash that's not available from the public processed crash API. That means you need an API token with permission to view personally identifiable information to see JIT classification. If you don't have an API token, then JIT-related crashes will end up with a different signature when running the signature generation command line interface in a local development environment. There's nothing we can do about this.
It's still a sad story for external users, but a step in the right direction. I didn't fully extract the library, so it's still hard to use for non-Socorro things. I did document some things we didn't previously know about signature generation. For example, now we know exactly what data it requires. We'll need to do more to make this viable for external uses.
The API is lacking. The pseudo-extracted library has a terrible API making it hard to use. We'll need to do more to make this viable for external uses.
Reducing scope helps things ship. Socorro has two developers  and in 2017q3, we were juggling a bunch of big projects. I ended up prioritizing and then deprioritizing this project several times. We had merge conflicts due to other work, so I rewrote it a few times. It would have been easier if there was less going on and if I had focused on this until it shipped. Life is hard! If I hadn't reduced the scope, it wouldn't have shipped in this environment.
|||Two Socorro developers for a codebase that's roughly 62k LOC Python and then 20k bash, SQL, JS, and other things.|
Conclusion and where we could go from here
I've been able to work on signature generation bugs I wasn't able to work on before. I can review signature generation pull requests and tell the author exactly how the change affects signatures. Theoretically, other people can test their changes now, though I don't know if anyone has.
In extracting signature generation, we improved the documentation for it and how it works as well as what parts of the raw and processed crash it requires.
It's easier to pull this code out of Socorro and use it elsewhere, thought probably still not viable.
I think all-in-all it's a good step in the right direction. But there's more we could do!
We could improve the API. The API for the pipeline is terrible. This is in part because I wanted to minimize the number of changes I made to extract it from being part of the processor. The signature generation pipeline doesn't need to think about raw and processed crash structures--it can just think about the data it's being passed in.
Similarly, it modifies the signature value in the processed crash. Instead of modifying the processed crash, it should be returning a signature structure which gets passed to the next rule in the pipeline. Then we can treat the crash data as immutable and reduce weirdo side-effects of the pipeline.
Improving the API will make it easier for others to use.
We could document all the bits that go into making a signature. The command line interface code extracts the bits it needs from the raw and processed crash, but no where do we explain what those bits are or what they look like. Fixing this would make it easier to use.
Further, it's likely we could reduce the data we need even more. Instead of providing all the threads of data, we only need the important one which is the crashing thread or thread 0 (I think that's what it is). We don't need the others. Maybe there's other data like that which we could remove?
We could extract into a standalone library or put it in a web service or something. In its current state, it's hard for anyone else to use. I was thinking we could turn it into a standalone library, but that has some tough maintenance costs around using it in Socorro and releases and such. There are solutions to those problems, but all of them involve an additional maintenance burden for a very very small team.
Ted suggested making a web service. That's definitely doable. Then anyone could use it regardless of the programming language they're using. It's worth asking around and seeing what the usage requirements and impact of doing this would be. Does it help some team a lot? Is it a silver bullet for work on Firefox stability?
The next step would probably involve talking with other groups who need to correlate crash data with Socorro and figuring out how can we do that in a way that maximizes ease of use but minimizes maintenance.
If you're in such a group, get a hold of me!
We could experimenting with other signature generation algorithms. Backtrace uses a machine learning algorithm on a large crash data set to suss out what makes a good signature and then uses that. They update their corpus periodically. We've probably got enough data that this could be an interesting approach. Maybe use it as a second categorization system?
There's probably a need for generation signatures using less information--maybe just using the stack. Does that help correlating with other data sets? Should crashes have multiple signatures?
That's the story of the signature generation command line interface. Any factual errors are because of a lack of ice cream while writing this.
If you have any questions or bump into bugs, we hang out on #breakpad on irc.mozilla.org. You can also write up a bug for Socorro.
Hopefully this helps. If not, let us know!