Supergroups

text/XXXX-supergroups.md

@@ -0,0 +1,275 @@
+* Start Date: 2022-06-22
+* RFC Type: feature / decision / informational
+* RFC PR: https://github.com/getsentry/rfcs/pull/29
+
+# Summary
+
+This RFC describes the motivation for supergroups.
+
+## Motivation
+
+When an error or exception is reported, it's desireable to fingerprint them so that the
+frequency of their ocurrence can be determined.  However what makes an error "the same"
+as another error?  Sentry has always run an algorithm to calculate the unique fingerprint
+of the error.  Any further error with the same fingerprint is fed into the same group and
+tallied up.  However no error is quite the same so the fingerprinting algorithm intentionally
+removed quite a bit of information or cleaned it up, to increase the chance of getting the
+same fingerprint.
+
+In doing so, there is a direct relationship to the creation of a Sentry issue to the
+calculation of the fingerprint.  This proposal wants to break this process up into two
+independent steps.
+
+It's evident from empiric evidence at Sentry that creating a overly precise fingerprint is
+problematic if each individual fingerprint creates a new issue in Sentry.  Because of this
+coupling the grouping algorithms have learned a few tricks over the years to disregard
+certain patterns that are known to be unstable between errors.  However there are a few
+fundamental issues that the grouping algorithm cannot address today.
+
+### Non Stack Fingerprints
+
+The primary way of creating fingerprints for errors are components of the stack trace.  There
+are however situations where a stack trace is unavailable or the stack trace does not have
+enough useful information (ie: the entire stack is discarded).  In that case Sentry needs to
+fall back to alternative grouping information which is often a string supplied with the error
+(the error or log message).  These typically contain formatted strings which are different
+between error instances.  The classical example is an error such as `"Worker process crashed (PID=23)"`.
+
+Here even multiple years of investment into the grouping algorithm has not resulted in good
+fingerprints.  We generally undergroup compared to what a human would expect, however the
+approach taken today is limited as it needs to generate a unique fingerprint.
+
+### Caller vs Callee Stacks
+
+One challenge with grouping based on stacks is that bugs might be happening in functions which
+are invoked from different places in the codebase.  Imagine a function that returns the currently
+authenticated user (`get_current_user`).  If this function is used all across the code base
+in different places, then a newly introduced regression can cause a range of different stack
+traces to be generated which in turn results in independent fingerprints.  To make this more
+concrete imagine the function fails with an `AttributeError` and is invoked in the following
+ways:
+
+```
+Traceback (most recent call last):
+  File "demo.py" in login
+  File "demo.py" in validate_access
+  File "demo.py" in get_current_user
+AttributeError: 'NoneType' object has no attribute 'username'
+```
+
+vs
+
+```
+Traceback (most recent call last):
+  File "demo.py" in checkout
+  File "demo.py" in process_transaction
+  File "demo.py" in get_current_user
+AttributeError: 'NoneType' object has no attribute 'username'
+```
+
+Here it seems obvious in a way that these are the same error beacuse the issue
+is clearly in the `get_current_user` function.  Yet the two stack traces are
+different (`login | validate_access | get_current_user` and
+`checkout | process_transaction | get_current_user`).  In this case we would say that the error
+is in the callee and would expect them to be grouped the same.
+
+On the other hand there can be exactly the same type of stack trace situation, but instead the
+error is in fact in the caller due to bad input.  Take for instance a function which parses a
+UUID (`parse_uuid`).  As part of the contract of this function we expect that the function will
+raise an error on bad input (`InvalidUuid`).  Now this function ends up being called in two
+different places:
+
+```
+Traceback (most recent call last):
+  File "demo.py" in load_images
+  File "demo.py" in parse_uuid
+InvalidUuid
+```
+
+vs
+
+```
+Traceback (most recent call last):
+  File "demo.py" in parse_request
+  File "demo.py" in parse_uuid
+InvalidUuid
+```
+
+Here the error is in fact not in `parse_uuid`.  The function operates as expected, but the
+calling functions are failing to catch down the invalid input error.  We at this point do not
+even know yet if the `parse_request` or `load_images` function are supposed to catch down the
+error or if someone higher up in the stack is supposed to, but we do know that it somewhere made
+it to a point where a Sentry error was created.  In this case we would not want this to group
+together under `parse_uuid`.
+
+### Environmental Failures
+
+Another class of errors that causes challenges today are what we call "environmental failures".  These
+are failures that happen because something in the environment degraded and is starting to fail
+unexpectedly.  A class of these environmental failures are network problems.  If for instance your database
+becomes unavailable, then a range of functions all over the product will start to fail in yet unseen
+ways.  This will create a lot of new fingerprints and as a result will create a number of new Sentry
+errors.  Particularly during incidents it's very common that a Sentry project will generate a huge
+number of issues that are unlikely to happen a second time and your only solution is to mass resolve
+these.
+
+Some of these environmental failures could be clustered from a fingerprinting point of view at the
+level of just the exception.  For instance one could make a relatively drastic call and decide that
+any `SocketError` shares the same fingerprint.  This however is unscalable for large software bases
+where different parts of the code base connect to different endpoints and as a user you want to be
+able to distinguish between these.
+
+### Snapshot or Sampled Stack Traces
+
+When a error is reported to Sentry in many cases the stack trace is coming directly from the exception
+object.  This means that when the error is created, it also carries at least an approximation of the
+call stack.  There are however situations where errors are reported to Sentry and the stack trace is
+fabricated.  Many of these are from situations where an error was not actually happening, but the
+SDK decided to report one anyways from a watchdog.  The most common example is something like a
+deadlock or hang detection.  If a certain task is not making progress for an extended period of time
+SDKs will often monitor what is happening from a background thread and issue a hang warning as Sentry
+error.  In that case the SDK goes in and creates a stack trace for another thread which might actually
+be busy doing something instead of hanging on a lock.  In that case the stack is highly unstable.
+
+This is a common issue with ANR (Application Not Responding) reports on mobile for instance where we
+already have to opt to a different grouping algorithm altogether as it creates too much noise otherwise.
+
+### Runtime Unstable Stacks
+
+The fingerprinting algorithm currently replies heavily on the fact that the same error produces the
+same fingerprint.  Unfortunatley there are limitations in runtimes that makes this impossible at times.
+For instance when working with browser javascript engines, the stack traces between different browsers
+are never quite the same.
+
+Take for instance this little JavaScript example:
+
+```javascript
+function failingFunction() {
+  throw new Error("broken");
+}
+
+const result = [1, 2, 3].map((item) => {
+  if (item % 2 == 1) {
+    failingFunction();
+  }
+  return item;
+});
+```
+
+In Firefox the stack trace looks roughly like this:
+
+```
+failingFunction
+result<
+(anonymous)
+(anonymous)
+```
+
+Compare this to Chrome:
+
+```
+failingFunction
+(anonymous)
+Array.map
+(anonymous)
+```
+
+Or Safari:
+
+```
+failingFunction
+(anonymous)
+map@[native code]
+(global code)
+```
+
+Or node:
+
+```
+failingFunction
+(anonymous)
+Array.map
+Object.<anonymous>
+Module._compile
+Object.Module._extensions..js
+Module.load
+Function.Module._load
+Function.executeUserEntryPoint
+```
+
+There is little agreement on what the stack should look like and it cannot even be properly
+normalized.  Sure there is something that can be detected or cleaned up, but fundamentally the
+stack traces are differences are big enough that generating the complete same fingerprint is
+close to impossible and definitely impossible in the general case.
+
+### Compiler Unstable Stacks
+
+Similar to the situation where the runtime creates difference stack traces, there are also
+issues where compilers or transpilers are involved.  The tricky case here are largely compiler
+optimizations.  The most trivial example are inlines.  If the compiler choses to inline a frame
+the stack traces in the trivial case changes.  This would not appear to be much of an issue as
+the debug information carries enough information to know where a function was inlined and that
+can be used to correct the stack trace.  However unfortunately limitations in the debug information
+format often cause inline frames to look slightly different than regular frames.
+
+The reason for this issue is that for instance in C++ and many other languages there are really
+two different ways to describe a function.  One is the mangled format and one is the demangled
+format.  Neither of these formats are particularly stable and there are no agreed upon standards.
+In PDB for instance for mangled functions there is no general way to retrieve either format for
+functions.  Many inline functions do not carry the same amount of information as a non inlined one.
+For instance take this function:
+
+```
+template <class T> const T &min(const T &a, const T &b)
+{
+    return !(b < a) ? a : b;
+}
+```
+
+When invoked with an `int` the function when not inlined is typically recoded as
+`module::min<int>(const int &, const int &)` or `min(const int &, const int &)`
+in the stack trace.  However when inlined the function on Windows platforms
+might only be called `min`.  However that's not true in the general case either.
+Many functions do retain their types, some do not.
+
+This problem becomes even larger when dealing with compiler generated code.  For instance C++
+knows a lot of different destructors and they can all look very different and they will look
+different depending on if they are inlined or not.
+
+As such from one release to another the fingerprint generated for a stack might change entirely.
+
+### Platform Unstable Stacks
+
+Related to the former case another complication arises from platform provided primitives that might
+end up in the stack trace.  Imagine for instance an application written in Rust that runs on macOS,
+Windows and Linux and has a trivial user introduced panic in a function invoked from an event loop.
+The event loop part of the stack will look different on every platform, yet the user code on top
+will look the very same.  On Linux it might use some epoll abstraction, in Windows IOCP or kqueue on
+mac.  Similar problems happy with mutexes, thread spawning code or more.
+
+None of the platform differences are relevant to the user bug, yet the same issue will create three
+different fingerprints, one per platform.
+
+We can also not universally exclude platform specific code because there might very well be a bug
+in this threading code.  For instance you would not want to have all the event handling code be
+excluded from stack traces for the case that the bug is in fact in the platform specific event
+handling code.
+
+## Proposed Changes
+
+The relationship of event to issue is very strong in Sentry.  There are directly workflows associated
+with it and there are quite significant technical hurdles to this.  The proposal is to not touch this
+association but instead augment it on a higher level.  The idea is to actually create these Sentry
+issues still based on the unique fingerprint but to lower the consequences of such a group being
+created.  Related fingerprints could then at regular intervals be sweeped up in to larger super
+groups.
+
+This means that from a user's point of view the individual groups still exist, but the main way to
+interact with them would be larger groups that group around the already existing smaller groups.
+Since the events stay associated with their fingerprint created group the event itself is immutable
+and stays that way, but the merge just becomes an association of that group with the supergroup.
+
+The way the individual fingerprints are associated with the supergroup is not defined by this
+proposal.  The idea is to create the possibility and to then experiment with ways in which this
+can happen.
+