[heka] State and future of Heka

[Rob Miller]

Hi everyone,

I'm loooong overdue in sending out an update about the current state of 
and plans for Heka. Unfortunately, what I have to share here will 
probably be disappointing for many of you, and it might impact whether 
or not you want to continue using it, as all signs point to Heka getting 
less support and fewer updates moving forward.

The short version is that Heka has some design flaws that make it hard 
to incrementally improve it enough to meet the high throughput and 
reliability goals that we were hoping to achieve. While it would be 
possible to do a major overhaul of the code to resolve most of these 
issues, I don't have the personal bandwidth to do that work, since most 
of my time is consumed working on Mozilla's immediate data processing 
needs rather than general purpose tools these days. Hindsight 
(https://github.com/trink/hindsight), built around the same Lua sandbox 
technology as Heka, doesn't have these issues, and internally we're 
using it more and more instead of Heka, so there's no organizational 
imperative for me (or anyone else) to spend the time required to 
overhaul the Go code base.

Heka is still in use here, though, especially on our edge nodes, so it 
will see a bit more improvement and at least a couple more releases. 
Most notably, it's on my list to switch to using the most recent Lua 
sandbox code, which will move most of the protobuf processing to custom 
C code, and will likely improve performance as well as remove a lot of 
the problematic cgo code, which is what's currently keeping us from 
being able to upgrade to a recent Go version.

Beyond that, however, Heka's future is uncertain. The code that's there 
will still work, of course, but I may not be doing any further 
improvements, and my ability to keep up with support requests and PRs, 
already on the decline, will likely continue to wane.

So what are the options? If you're using a significant amount of Lua 
based functionality, you might consider transitioning to Hindsight. Any 
Lua code that works in Heka will work in Hindsight. Hindsight is a much 
leaner and more solid foundation. Hindsight has far fewer i/o plugins 
than Heka, though, so for many it won't be a simple transition.

Also, if there's someone out there (an organization, most likely) that 
has a strong interest in keeping Heka's codebase alive, through funding 
or coding contributions, I'd be happy to support that endeavor. Some 
restrictions apply, however; the work that needs to be done to improve 
Heka's foundation is not beginner level work, and my time to help is 
very limited, so I'm only willing to support folks who demonstrate that 
they are up to the task. Please contact me off-list if you or your 
organization is interested.

Anyone casually following along can probably stop reading here. Those of 
you interested in the gory details can read on to hear more about what 
the issues are and how they might be resolved.

First, I'll say that I think there's a lot that Heka got right. The 
basic composition of the pipeline (input -> split -> decode -> route -> 
process -> encode -> output) seems to hit a sweet spot for composability 
and reuse. The Lua sandbox, and especially the use of LPEG for text 
parsing and transformation, has proven to be extremely efficient and 
powerful; it's the most important and valuable part of the Heka stack. 
The routing infrastructure is efficient and solid. And, perhaps most 
importantly, Heka is useful; there are a lot of you out there using it 
to get work done.

There was one fundamental mistake made, however, which is that we 
shouldn't have used channels. There are many competing opinions about Go 
channels. I'm not going to get in to whether or not they're *ever* a 
good idea, but I will say unequivocally that their use as the means of 
pushing messages through the Heka pipeline was a mistake, for a number 
of reasons.

First, they don't perform well enough. While Heka performs many tasks 
faster than some other popular tools, we've consistently hit a 
throughput ceiling thanks to all of the synchronization that channels 
require. And this ceiling, sadly, is generally lower than is acceptable 
for the amount of data that we at Mozilla want to push through our 
aggregators single system.

Second, they make it very hard to prevent message loss. If unbuffered 
channels are used everywhere, performance plummets unacceptably due to 
context-switching costs. But using buffered channels means that many 
messages are in flight at a time, most of which are sitting in channels 
waiting to be processed. Keeping track of which messages have made it 
all the way through the pipeline requires complicated coordination 
between chunks of code that are conceptually quite far away from each other.

Third, the buffered channels mean that Heka consumes much more RAM than 
would be otherwise needed, since we have to pre-allocate a pool of 
messages. If the pool size is too small, then Heka becomes susceptible 
to deadlocks, with all of the available packs sitting in channel queues, 
unable to be processed because some plugin is blocked on waiting for an 
available pack. But cranking up the pool size causes Heka to use more 
memory, even when it's idle.

Hindsight avoids all of these problems by using disk queues instead of 
RAM buffers between all of the processing stages. It's a bit 
counterintuitive, but at high throughput performance is actually better 
than with RAM buffers, because a) there's no need for synchronization 
locks and b) the data is typically read quickly enough after it's 
written that it stays in the disk cache.

There's much less chance of message loss, because every plugin is 
holding on to only one message in memory at a time, while using a 
written-to-disk cursor file to track the current position in the disk 
buffer. If the plug is pulled mid-process, some messages that were 
already processed might be processed again, but nothing will be lost, 
and there's no need for complex coordination between different stages of 
the pipeline.

Finally, there's no need for a pool of messages. Each plugin is holding 
some small number of packs (possibly as few as one) in its own memory 
space, and those packs never escape that plugin's ownership. RAM usage 
doesn't grow, and pool exhaustion related deadlocks are a thing of the past.

For Heka to have a viable future, it would basically need to be updated 
to work almost exactly like Hindsight. First, all of the APIs would need 
to be changed to no longer refer to channels. (The fact that we exposed 
channels to the APIs is another mistake we made... it's now generally 
frowned upon in Go land to expose channels as part of your public APIs.) 
There's already a non-channel based API for filters and outputs, but 
most of the plugins haven't yet been updated to use the new API, which 
would need to happen.

Then the hard work would start; a major overhaul of Heka's internals, to 
switch from channel based message passing to disk queue based message 
passing. The work that's been done to support disk buffering for filters 
and outputs is useful, but not quite enough, because it's not scalable 
for each plugin to have its own queue; the number of open file 
descriptors would grow very quickly. Instead it would need to work like 
Hindsight, where there's one queue that all of the inputs write to, and 
another that filters write to. Each plugin reads through its specified 
input queue, looking for messages that match its message matcher, 
writing its location in the queue back to the shared cursors file.

There would also be some complexity in reconciling Heka's breakdown of 
the input stage into input/splitter/decoder with Hindsight's 
encapsulation of all of these stages into a single sandbox.

Ultimately I think this would be at least 2-3 months full time work for 
me. I'm not the fastest coder around, but I know where the bodies are 
buried, so I'd guess it would take anyone else at least as long, 
possibly longer if they're not already familiar with how everything is 
put together.

And that's about it. If you've gotten this far, thanks for reading. 
Also, thanks to everyone who's contributed to Heka in any way, be it by 
code, doc fixes, bug reports, or even just appreciation. I'm sorry for 
those of you using it regularly that there's not a more stable future.

Regards,

-r