The GitLab Rails application runs on Puma, a multi-threaded Rack application server written in the new Ruby. We recently updated Puma to major version 5, which introduced a number of important changes, including support for compaction, a technique to reduce memory fragmentation in the Ruby heap.
In this post we will describe what Puma's "nakayoshi fork" does, what compaction is, and some of the challenges we faced when first deploying it.
Nakayoshi: A friendlier
Puma 5 added a new configuration switch:
nakayoshi_fork. This switch affects Puma's behavior when
forking new workers from the primary process. It is largely based on a Ruby gem of the same name
but adds new functionality. More specifically, enabling
nakayoshi_fork in Puma will result in two additional
steps prior to forking into new workers:
Tenuring objects. By running several minor garbage collection cycles ahead of a
fork, Ruby can promote survivors from the young to the old generation (referred to as "tenuring"). These objects are often classes, modules, or long-lived constants that are unlikely to change. This process makes forking copy-on-write friendly because tagging an object as "old" implies a write to the underlying heap page. Doing this prior to forking means the OS won't have to copy this page from the parent to the worker process later. We won't be discussing copy-on-write in detail but this blog post offers a good introduction to the topic and how it relates to Ruby and pre-fork servers.
Heap compaction. Ruby 2.7 added a new method
GC.compact, which will reorganize the Ruby heap to pack objects closer together when invoked.
GC.compactreduces Ruby heap fragmentation and potentially frees up Ruby heap pages so that the physical memory consumed can be reclaimed by the OS. This step only happens when
GC.compactis available in the version of Ruby that is in use (for MRI, 2.7 or newer).
In the remainder of this post, we will look at:
GC.compactworks and its potential benefits.
- Why using C-extensions can be problematic when using compaction.
- How we resolved a production incident that crashed GitLab.
- What to look out for before enabling compaction in your app, via
How compacting garbage collection works
The primary goal of a compacting garbage collector (GC) is to use allocated memory more effectively, which increases the likelihood of the application using less memory over time. Compaction is especially important when processes can share memory, as is the case with Ruby pre-fork servers such as Puma or Unicorn. But how does Ruby accomplish this?
Ruby manages its own object heap by allocating chunks of memory from the operating system called pages (a confusing term since Ruby heap pages are distinct from the smaller memory pages managed by the OS itself). When an application asks to create a new object, Ruby will try to find a free object slot in one of these pages and fill it. As objects are allocated and deallocated over the lifetime of the application, this can lead to fragmentation, with pages being neither entirely full nor entirely empty. This is the primary cause for Ruby's infamous runaway memory problem: Since the available space isn't optimally used, pages will rarely be entirely empty and become "tomb pages" which means it is necessary for the pages to be empty for them to be deallocated.
Ruby 2.7 added a new method,
GC.compact, which aims to address this problem by walking the entire
Ruby heap space and moving objects around to obtain tightly packed pages. This process will ideally make
some pages unused, and unused memory can be reclaimed by the OS. Watch this video from RubyConf 2019 where Aaron Patterson, the author of this feature, gave a good introduction to compacting GC.
Compaction is a fairly expensive task since Ruby needs to stop-the-world for a complete heap reorganization so
its best to perform this task before forking a new worker process, which is why Puma 5 included this step when performing
nakayoshi_fork. Moreover, running compaction before forking
into worker processes increases the chance of workers being able to share memory.
We were eager to enable this feature on GitLab to see if it would reduce memory consumption, but things didn't entirely go as planned.
Inside the incident
After extensive testing via our automated performance test suite and in preproduction environments, we felt ready to explore compaction on production nodes. We kept a detailed, public record of what happened during this production incident, but the key details are summarized below:
- The deployment passed the canary stage, meaning workers who had their heaps compacted were serving traffic successfully at this point.
- Sometime during the full fleet rollout, problems emerged: Error rates started spiking but not across the entire fleet. This phenomenon is odd because errors tend to spread across all workers due to load balancing.
- The error messages surfacing in Sentry were mysterious at best:
ActionView::Template::Error uninitialized constant #<Class:#GrapePathHelpers::DecoratedRoute:0x00007f95f10ea5b8>::UNDERSCORE. Remember this error message for later.
- We discovered the affected workers were segfaulting in
hamlit, a high-performance HAML compiler. Hamlit uses a C-extension to achieve better performance. The segfaulting and the fact that we were rolling out an optimization that touches GC-internal structures was a tell-tale sign that compaction was likely to be the cause.
- We rolled back the change to quickly recover from the outage.
How we diagnosed the problem
We were disappointed by this setback and wanted to understand why the outage occurred. Fortunately, Ruby provides detailed stack traces when crashing in C-extensions. The most effective way to quickly analyze these is to look for transitions where a C-extension calls into the Ruby VM or vice versa. These lines therefore caught our attention:
... /opt/gitlab/embedded/lib/libruby.so.2.7(sigsegv+0x52) [0x7f9601adb932] signal.c:946 /lib/x86_64-linux-gnu/libc.so.6(0x7f960154c4c0) [0x7f960154c4c0] /opt/gitlab/embedded/lib/libruby.so.2.7(rb_id_table_lookup+0x1) [0x7f9601b15e11] id_table.c:227 /opt/gitlab/embedded/lib/libruby.so.2.7(rb_const_lookup+0x1e) [0x7f9601b4861e] variable.c:3357 /opt/gitlab/embedded/lib/libruby.so.2.7(rb_const_get+0x39) [0x7f9601b4a049] variable.c:2339 # ^--- Ruby VM functions /opt/gitlab/embedded/lib/ruby/gems/2.7.0/gems/hamlit-2.11.0/lib/hamlit/hamlit.so(str_underscore+0x16) [0x7f95ee3518f8] hamlit.c:17 /opt/gitlab/embedded/lib/ruby/gems/2.7.0/gems/hamlit-2.11.0/lib/hamlit/hamlit.so(rb_hamlit_build_id) hamlit.c:100 # ^-- hamlit C-extension ...
The topmost stack frame reveals the preceeding calls led to a segmentation fault (
We highlighted the lines where Hamlit calls back into Ruby: In a function called
was called by
rb_* prefix tells us that this is a C-function we can call from Ruby,
and indeed it is used by
Hamlit::AttributeBuilder to construct DOM
But we still don't know why it is crashing. Next, we need to inspect what happens in
We can see that this function performs a constant lookup on
mAttributeBuilder – searching
for a constant called
UNDERSCORE. When following the breadcrumbs it turns out to simply be the string
It is this lookup that failed.
UNDERSCORE? That sounds familiar. Recall the top-level error messages:
ActionView::Template::Error uninitialized constant #<Class:#GrapePathHelpers::DecoratedRoute:0x00007f95f10ea5b8>::UNDERSCORE
GrapePathHelpers is clearly not a Hamlit class. Hamlit is trying to look up its own
constant on a class in the
grape gem, an entirely different library
that is not involved in HTML rendering at all and there is no such constant defined on Grape's
DecoratedRoute class either.
Now the penny dropped – remember how compaction moves around objects in Ruby's heap space? Classes in
Ruby are objects too, so
GC.compact must have moved a Grape class into an object slot that was previously
occupied by a Hamlit class object, but Hamlit's C-extension never saw it coming!
How we solved the problem
To be clear, what happened above should not happen with a well-behaved C-extension. Compaction was developed carefully with support for C-extensions that predate Ruby 2.7, so all existing Ruby gems would continue to operate normally.
So what went wrong? When a C-extension allocates Ruby objects, it must mark them for as long as they are alive. A marked object will not be garbage collected and because the Ruby GC cannot reason about objects outside of its own purview (i.e., objects created from Ruby code), it needs to rely on C-extensions to correctly mark and unmark objects themselves.
Now comes the twist: Marked objects can be moved during compaction and existing C-extensions can't cope with an object they hold pointers to suddenly move into a different slot. Therefore, Ruby 2.7 does something clever: It "pins" objects allocated with the mark function that existed prior to Ruby 2.7, meaning the pinned objects are not allowed to move during compaction. For new code, it introduces a special mark-but-don't-pin function that will also allow an object to move, giving gem authors the opportunity to make their libraries compaction-aware.
Hamlit does not implement compaction support, so this could only mean one thing:
Hamlit wasn't even properly marking those objects, otherwise Ruby 2.7
would have automatically pinned them so they wouldn't move during compaction.
After discussing an attempted fix we submitted but without
a reliable way to reproduce the issue for everyone, the Hamlit author decided to sidestep the
problem by resolving those constants statically instead
and marking each via
This change landed in Hamlit 2.14.2
which we confirmed resolves the issue.
The next steps
It is exciting to see that the Ruby community is making progress on making Ruby a more memory-efficient language but we learned that we need to step carefully when introducing such wide-reaching changes to a large application like GitLab. It is difficult to investigate and fix problems that crash the Ruby VM, which is more likely for any library that uses C-extensions.
Two particular action items we took away from this were:
- More reliable detection of compaction-related issues in CI. We're not going to sugar-coat this:
We detected the problem late. Our comprehensive test suite was passing, our QA and performance tests
on staging environments passed, and the problem didn't even show up in canary deployments. Ideally, we
would have caught this issue with our automated test suite. One way to test whether compaction causes problems
is by using
GC.verify_compaction_references– this is a rather crude tool because it requires keeping two copies of the Ruby heap, which can be prohibitively expensive in terms of memory use. We have therefore not yet decided how to approach this.
- Improve our ability to roll out system configuration gradually. Puma is part of our core infrastructure, since it sits in the path of every web request, which makes it especially risky to experiment with Puma configuration. GitLab already supports feature flags to allow developers to roll out product changes gradually, but it presents us with a catch-22 when making changes at the infrastructure level, because to query the state of a feature flag, the infrastructure needs to already be up and running. It would be ideal to have a similar mechanism for system configuration, which we are currently exploring.
While performance is a major focus for us at the moment it must not compromise availability. We will continue to monitor developments in the Ruby community around compaction support, but decided to not use it in production at this point in time since the gains don't appear to outweigh the risks.