Socorro: Schema based overhaul of crash ingestion: retrospective (2022)

Project

time:

2+ years

impact:

• radically reduced risk of data leaks due to misconfigured permissions

• centralized and simplified configuration and management of fields

• normalization and validation performed during processing

• documentation of data reviews, data caveats, etc

• reduced risk of bugs when adding new fields--testing is done in CI

• new crash reporting data dictionary with Markdown-formatted descriptions, real examples, relevant links

Summary

I've been working on Socorro (crash ingestion pipeline at Mozilla) since the beginning of 2016. During that time, I've focused on streamlining maintainence of the project, paying down technical debt, reducing risk, and improving crash analysis tooling.

One of the things I identified early on is how the crash ingestion pipeline was chaotic, difficult to reason about, and difficult to document. What did the incoming data look like? What did the processed data look like? Was it valid? Which fields were protected? Which fields were public? How do we add support for a new crash annotation? This was problematic for our ops staff, engineering staff, and all the people who used Socorro. It was something in the back of my mind for a while, but I didn't have any good thoughts.

In 2020, Socorro moved into the Data Org which has multiple data pipelines. After spending some time looking at how their pipelines work, I wanted to rework crash ingestion.

The end result of this project is that:

1. the project is easier to maintain:

   • adding support for new crash annotations is done in a couple of schema files and possibly a processor rule

2. risk of security issues and data breaches is lower:

   • typos, bugs, and mistakes when adding support for a new crash annotation are caught in CI

   • permissions are specified in a central location, changing permission for fields is trivial and takes effect in the next deploy, setting permissions supports complex data structures in easy-to-reason-about ways, and mistakes are caught in CI

3. the data is easier to use and reason about:

   • normalization and validation of crash annotation data happens during processing and downstream uses of the data can expect it to be valid; further we get a signal when the data isn't valid which can indicate product bugs

   • schemas describing incoming and processed data

   • crash reporting data dictionary documenting incoming data fields, processed data fields, descriptions, sources, data gotchas, examples, and permissions

What is Socorro?

Socorro is the crash ingestion pipeline for Mozilla products like Firefox, Fenix, Thunderbird, and MozillaVPN.

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 submitted crash report, processes it, and has tools for viewing and analyzing crash data.

State of crash ingestion at the beginning

The crash ingestion system was working and it was usable, but it was in a bad state.

• Poor data management

  Normalization and validation of data was all over the codebase and not consistent:

  • processor rule code

  • AWS S3 crash storage code

  • Elasticsearch indexing code

  • Telemetry crash storage code

  • Super Search querying and result rendering code

  • report view and template code

  • signature report code and template code

  • crontabber job code

  • any scripts that used the data

  • tests -- many of which had bad test data so who knows what they were really testing

  Naive handling of minidump stackwalker output which meant that any changes in the stackwalker output were predominantly unnoticed and there was no indication as to whether changed output created issues in the system.

  Further, since it was all over the place, there were no guarantees for data validity when downloading it using the RawCrash, ProcessedCrash, and SuperSearch APIs. Anyone writing downstream systems would also have to normalize and validate the data.

• Poor permissions management

  Permissions were defined in multiple places:

  • Elasticsearch json redactor

  • Super Search fields

  • RawCrash API allow list

  • ProcessedCrash API allow list

  • report view and template code

  • Telemetry crash storage code

  • and other places

  We couldn't effectively manage permissions of fields in the stackwalker output because we had no idea what was there.

• Poor documentation

  No documentation of crash annotation fields other than CrashAnnotations.yaml which didn't enforce anything in crash ingestion (process, valid type, data correctness, etc) and was missing important information like data gotchas, data review urls, and examples.

  No documentation of processed crash fields at all.

• Making changes was high risk

  Changing fields from public to protected was high risk because you had to find all the places it might show up which was intractable. Adding support for new fields often took multiple passes over several weeks because we'd miss things. Server errors happend with some regularity due to weirdness with crash annotation values affecting the Crash Stats site.

• Tangled concerns across the codebase

  Lots of tangled concerns where things defined in one place affected other places that shouldn't be related. For example, the Super Search fields definition was acting as a "schema" for other parts of the system that had nothing to do with Elasticsearch or Super Search.

• Difficult to maintain

  It was difficult to support new products.

  It was difficult to debug issues in crash ingestion and crash reporting.

  The Crash Stats webapp contained lots of if/then/else bits to handle weirdness in the crash annotation values. Nulls, incorrect types, different structures, etc.

  Socorro contained lots of vestigial code from half-done field removal, deprecated fields, fields that were removed from crash reports, etc. These vestigial bits were all over the code base. Discovering and removing these bits was time consuming and error prone.

  The code for exporting data to Telemetry built the export data using a list of fields to exclude rather than a list of fields to include. This is backwards and impossible to maintain--we never should have been doing this. Further, it pulled data from the raw crash which we had no validation guarantees for which would cause issues downstream in the Telemetry import code.

  There was no way to validate the data used in the unit tests which meant that a lot of it was invalid. We had no way to validate the test data which meant that CI would pass, but we'd see errors in our stage and production environments.

• Different from other similar systems

  In 2020, Socorro was moved to the Data Org in Mozilla which had a set of standards and conventions for collecting, storing, analyzing, and providing access to data. Socorro didn't follow any of it which made it difficult to work on, to connect with, and to staff. Things Data Org has that Socorro didn't:

  • a schema covering specifying fields, types, and documentation

  • data flow documentation

  • data review policy, process, and artifacts for data being collected and how to add new data

  • data dictionary for fields for users including documentation, data review urls, data gotchas

In summary, we had a system that took a lot of effort to maintain, wasn't serving our users' needs, and was high risk of security/data breach.

Project plan

Many of these issues can be alleviated and reduced by moving to a schema-driven system where we:

1. define a schema for annotations and a schema for the processed crash

2. change crash ingestion and the Crash Stats site to use those schemas

When designing this schema-driven system, we should be thinking about:

1. how easy is it to maintain the system?

2. how easy is it to explain?

3. how flexible is it for solving other kinds of problems in the future?

4. what kinds of errors will likely happen when maintaining the system and how can we avert them in CI?

5. what kinds of errors can happen and how much risk do they pose for data leaks? what of those can we avert in CI?

6. how flexible is the system which needs to support multiple products potentially with different needs?

I worked out a minimal version of that vision that we could migrate to and then work with going forward.

The crash annotations schema should define:

1. what annotations are in the crash report?

2. which permissions are required to view a field

3. field documentation (provenance, description, data review, related bugs, gotchas, analysis tips, etc)

The processed crash schema should define:

1. what's in the processed crash?

2. which permissions are required to view a field

3. field documentation (provenance, description, related bugs, gotchas, analysis tips, etc)

Then we make the following changes to the system:

1. write a processor rule to copy, nomralize, and validate data from the raw crash based on the processed crash schema

2. switch the Telemetry export code to using the processed crash for data to export

3. switch the Telemetry export code to using the processed crash schema for permissions

4. switch Super Search to using the processed crash for data to index

5. switch Super Search to using the processed crash schema for documentation and permissions

6. switch Crash Stats site to using the processed crash for data to render

7. switch Crash Stats site to using the processed crash schema for documentation and permissions

8. switch the RawCrash, ProcessedCrash, and SuperSearch APIs to using the crash annotations and processed crash schemas for documentation and permissions

After doing that, we have:

1. field documentation is managed in the schemas

2. permissions are managed in the schemas

3. data is normalized and validated once in the processor and everything uses the processed crash data for indexing, searching, and rendering

4. adding support for new fields and changing existing fields is easier and problems are caught in CI

Implementation decisions

Use JSON Schema.

Data Org at Mozilla uses JSON Schema for schema specification. The schema is written using YAML.

https://mozilla.github.io/glean_parser/metrics-yaml.html

The metrics schema is used to define metrics.yaml files which specify the metrics being emitted and collected.

For example:

https://searchfox.org/mozilla-central/source/toolkit/mozapps/update/metrics.yaml

One long long long term goal for Socorro is to unify standards and practices with the Data Ingestion system. Towards that goal, it's prudent to build out a crash annotation and processed crash schemas using whatever we can take from the equivalent metrics schemas.

We'll additionally need to build out tooling for verifying, validating, and testing schema modifications to make ongoing maintenance easier.

Use schemas to define and drive everything.

We've got permissions, structures, normalization, validation, definition, documentation, and several other things related to the data and how it's used throughout crash ingestion spread out across the codebase.

Instead of that, let's pull it all together into a single schema and change the system to be driven from this schema.

The schema will include:

1. structure specification

2. documentation including data gotchas, examples, and implementation details

3. permissions

4. processing instructions

We'll have a schema for supported annotations and a schema for the processed crash.

We'll rewrite existing parts of crash ingestion to use the schema:

1. processing

   1. use processing instructions to validate and normalize annotation data

2. super search

   1. field documentation

   2. permissions

   3. remove all the normalization and validation code from indexing

3. crash stats

   1. field documentation

   2. permissions

   3. remove all the normalization and validation code from page rendering

Only use processed crash data for indexing and analysis.

The indexing system has its own normalization and validation code since it pulls data to be indexed from the raw crash.

The crash stats code has its own normalization and validation code since it renders data from the raw crash in various parts of the site.

We're going to change this so that all normalization and validation happens during processing, the results are stored in the processed crash, and indexing, searching, and crash analysis only work on processed crash data.

By default, all data is protected.

By default, all data is protected unless it is explicitly marked as public. This has some consequences for the code:

1. any data not specified in a schema is treated as protected

2. all schema fields need to specify permissions for that field

3. any data in a schema is either:

   • marked public, OR

   • lists the permissions required to view that data

4. for nested structures, any child field that is public has public ancesters

We can catch some of these issues in CI and need to write tests to verify them.

This is slightly awkward when maintaining the schema because it would be more reasonable to have "no permissions required" mean that the field is public. However, it's possible to accidentally not specify the permissions and we don't want to be in that situation. Thus, we decided to go with explicitly marking public fields as public.

Work done

Random thoughts

This was a very very long-term project with many small steps and some really big ones. Getting large projects done is futile and the only way to do it successfully is to break it into a million small steps each of which stand on their own and don't create urgency for getting the next step done.

Any time I changed field names or types, I'd have to do a data migration. Data migrations take 6 months to do because I have to wait for existing data to expire from storage. On the one hand, it's a blessing I could do migrations at all--you can't do this with larger data sets or with data sets where the data doesn't expire without each migration becoming a huge project. On the other hand, it's hard to juggle being in the middle of multiple migrations and sometimes the contortions one has to perform are grueling.

If you're working on a big project that's going to require changing data structures, figure out how to do migrations early with as little work as possible and use that process as often as you can.

Conclusion and where we could go from here

This was such a huge project that spanned years. It's so hard to finish projects like this because the landscape for the project is constantly changing. Meanwhile, being mid-project has its own set of complexities and hardships.

I'm glad I tackled it and I'm glad it's mostly done. There are some minor things to do, still, but this new schema-driven system has a lot going for it. Adding support for new crash annotations is much easier, less risky, and takes less time.

It took me about a month to pull this post together.

That's it!

That's the story of the schema-based overhaul of crash ingestion. There's probably some bits missing and/or wrong, but the gist of it is here.

If you have any questions or bump into bugs, I hang out on #crashreporting on chat.mozilla.org. You can also write up a bug for Socorro.

Hopefully this helps. If not, let us know!