Hashie Breakup - Major Perf boost #773
Hashie, we've had some great times...but I think we need to break up. Your ties to omniauth are expensive, you're too high maintenance. We needed to take it to the next level but things aren't moving fast enough. It's just not working out...
I benched a Rails request with memory_profiler and found that Hashie allocates a TON of objects on each request:
This is one, request! I was curious if it was possible to remove hashie from Omniauth, and if so, would the result be better, faster, and use less memory. This PR does just that.
This patch removes Hashie as a dependency and replaces its usage with a custom data structure
It uses less memory:
Our structure uses
Our patch improves speed by:
Most tests pass with no modification and most interfaces remain unchanged. The biggest change is that there are no more predicate methods so use
There was one test that needed a
The case for removing hashie is pretty compelling, let's or discuss implementation. Afterwards we can talk about a plan to move omniauth towards real honest to goodness PORO config objects.
At the very least this will require a major version bump, because people have built thousands of applications depending on OmniAuth's AuthHash etc. as a hash-like structure, not a struct. I personally use query methods in my own applications.
I'm not saying let's not merge this, I think it's probably fine, but authentication is not typically a code path that needs to be heavily optimized. A 5% performance gain is great and all, but a given user is authenticating what, maybe once a week? Once a month? I worry about introducing such sweeping API changes for a noticeable but perhaps not-super-important perf bump.
referenced this pull request
Oct 1, 2014
Even after #774 we're leaving a good amount of perf on the table. There's likely some other big easy-ish wins. We can look at what the middleware is doing when it is not being used (no callback or auth path). It's fine if we're slower on the paths omniauth is registered at, but for every other request, omniauth should basically be invisible from a perf perspective. Even with #744, hashie is still expensive.
I agree with the major version bump, and if we're going to do that we might as well use custom objects instead of struct or hash types. The real kicker is the dual nature of data access, that flexibility comes at a cost. If we did that, we could see even more drastic savings than this PR.
In my ideal world I would get the top 10 omniauth strategies, build object types that work for those cases out of the box. After that iterate over the next 10-20 strategies by pointing the omniauth dependency at a branch and running the tests, modifying the strategies so that they pass and put together a guide on how to update strategies (common problems/pitfalls).
That all being said, it seems more worthwhile to look at non-api breaking improvements. Only then consider a major version rev project.
Thanks for the look. Going to close this, as it was perhaps more of a proof of concept, and could potentially be used to base future custom config object work.
Since this was linked to @schneems' blog post that rips Hashie as a "horrible" choice that should never be used in production, I wanted to share one thought I had while the page was loading."
"I bet this PR has more code without Hashie than with Hashie."
There are different types of ways to measure code. Milliseconds to execute a method is one. Lines of code to implement a method is another. Time to implement a solution is even another!
When I've chosen to use Hashie, it was to lower the lines of code and speed up development. If I were to drop Hashie in those cases, I'd end up with something just like this... a PR with many more additions than subtractions. Sure, it's faster, but that performance boost is not free.
Thanks for reading the article. This PR should be left for technical discussion about this specific code or the intent. This isn't the place for general discussion. Maybe http://www.reddit.com/r/ruby/comments/2pkzec/hashie_considered_harmful_an_ode_to_hash_and/ would be a better forum. I'll be happy to reply but don't want to clog up this issue.
@schneems Well, you are right... I should be more specific to this pull request.
I'd feel worse about omniaith with this pull request than without. It may perform better, but I feel that it's taking on a chunk of technical debt. The form of this debt is OmniStruct, the new 70-line, untested class that inherits from a ruby base class (usually a big no-no, right?)
The class may be tested passively through the existing tests, and often that's fine when refactoring code. However, when the refactor takes the form of a large class with its own responsibilities, I think it becomes necessary to test it in isolation. In the form that OmniStruct is in currently, I would not trust that it is fully covered. In a class with revision, I'd consider this unsafe.
The missing tests are even more important when considering the fact that hashie's business is not omniauth's business. I bet hashie has tests around the fine details of these behaviors because that's why hashie exists. Omniauth's focus is on its own business. Omniauth's needs only a subset of hashie behavior, but any detail that it pulls from hashie should be managed with extreme caution. I don't see that caution in this code, and I especially don't see it in the author's comments. I see someone who is focused on a tiny bit of performance and personal values, with no mention of the costs.
Choices in programming are usually a trade-off. You get something, but you give something away. A reasoned approach to changes involved acknowledging both sides. I raise a red flag when I see the person making the changes focused on one.
I know this PR has been closed, but moving forward I implore the authors to treat it seriously.
Responses to your comments in turn.
Agreed, that's why we closed it. This PR didn't go far enough. It tried to support the current interface, in its (almost) entirety when the right way to do it would be to introduce proper config with real honest to goodness objects and methods. I talked about this #773 (comment)
The current config object being based on Hashie has TREMENDOUS technical debt, which is why this PR was so difficult and why the
Said this earlier "we might as well use custom objects instead of struct or hash types"
Nope. This PR has 112 lines added and 32 lines deleted (net of 80 lines added). Removing the Hashie dependency (only mash) is around 213 lines (without comments). Hashie depends on
Now i know what you're thinking, hashie is a dependency, we don't get to count those lines. Well...when you use a third party library you become a maintainer of that code, Michael actually wrote the first version of Hashie and was quite literally the maintainer of the code. Even if that wasn't the case, if one of those 338 lines breaks something in omniauth, it's the maintainer of omniauth's job to find it, report it or fix it. It also increases the surface area that the maintainer needs to know.
Let's take a look. If you're using Hashie::Mash you need to know a bunch of methods:
And it doesn't include all those magic methods added by
In comparison, my proposed spike has less than half the methods (but the same method missing behavior). By introducing Hashie into your project you need to know more.
So we've got fewer lines of code to maintain and a much more defined interface (176 - 71 = 105 fewer methods).
I encourage people to use and share libraries but you're kidding yourself if you don't think you introduce technical debt into your project every dependency you add. When Rails 2 to Rails 3 broke a ton of gems and simultaneously a bunch of gem authors decided they didn't want to maintain those gems anymore. Guess what...every single dependency in your app became your responsibility. If it's code that is running on your server, it doesn't matter how it got there, it's your responsibility.
I completely agree here. When Omniauth passes your Rails app a config object and you accidentally call a method on it that doesn't exist and no exceptions get raised there is a very real cost to the time you will spend fixing your Rails code. An explicit interface would have made the problem so obvious it would have taken seconds to see there was an exception. If i could accurately measure the times i've misconfigured an omniauth strategy and spent hours debugging, I would gladly list them here.
I think what you're getting at, is that you don't want to maintain things like hashes with indifferent access and
I disagree. Omniauth must know about how hashie works to perform correctly. When you touch code that hashie touches, you must know how to use hashie. Omniauth uses hashie, therefore you must know how hashie works to work in the omniauth codebase.
The problem here is we pass this config object that is based on Hashie to 3rd party libraries and "strategies", even if it only needs a subset of the behavior, it is giving all the behavior to all the libraries. So while this code doesn't call a specific method, a third party library could. If we were going to actually go forwards, we would need to deprecate like crazy see my comment and example code here: zmoazeni/harvested#83 (comment)
The performance gains here are anything but tiny. Five percent of total rails application request speed isn't tiny, especially considering that Omniauth shouldn't be doing anything on those requests. I struggle for months on code in Rails internals to get just 1~2% faster on an overall request. 5% gain on one relatively small code change is HUGE. While i emphasized performance in the pull request, i should have talked about the biggest benefits or rather the "costs".
The "costs" here provide us with a cleaner interface, less total lines of code to maintain and fewer methods to have to understand due to removing Hashie as a dependency.
What i think you meant here was "You seem to be only focussing on the performance, how does this impact usability? Anything else this gives up?" The point of the article is that removing hashie makes your code faster, cleaner, and easier to use. You're giving up flexibility, but it's a false flexibility. The flexibility provided by Hashie also introduces problems.
Totally agree. What you might not realize is that it took me 16 hours to write this pull request and get all the tests working. I did it on my flight to and from RubyKaigi in Japan. This doesn't include all the time I spent writing the debugging and benchmarking tools. That was just to get the existing tests to pass. After nearly two straight (business) days of coding, i'm more interested in throwing the idea over the fence and seeing what comes of it than shooting for 100% on rcov. I can assure you that had anything merged into master I would have tested the crap out of it.
Telling a closed pull request it needs tests is not helpful.
You're right it's not free. It took me 16 hours to write that patch, and as we've mentioned it's not even the final solution. If hashie had never been introduced in the first place, then it would have been like getting a 5% performance boost for from the start. Did i mention we would also get a better interface and easier code to work with? Removing Hashie is a win, win, win. Because when I win and you win, we all win.
All other arguments aside, it comes down to this: 5% less time spent on every request to applications using Rack. I wouldn't be surprised, too, if more configuration-heavy strategies saw larger speed improvements, but that's conjecture.
This as an added benefit is great too.
This, or the discussed extension to scratch-building a configuration system, should pretty clearly (to me) be the way forward for Omniauth.
@schneems I won't keep this going forever, especially since it's closed. But a couple points:
You're equivocating. I pointed that this PR is built on a a large, complicated, untested class -- a class which wouldn't have to exist if Hashie was used. This is a fact, and a matter of basic math.
I've asserted that this fact, alone, shows that there is a cost to removing Hashie. Or to look at it from the other side, Hashie provides a value at a cost. I can't convey developer experience in writing, and I doubt I have to with you... but good programming requires judging the costs/benefits of multiple options, and then making the best decision.
I'm raising my red flag because I don't see any acknowledgment of this cost/benefit. Hey, I have to do this with myself all the time. I'll start on a new feature or "enhancement" for all sorts of reasons... but honestly, deep-in-my-heart, I know that it's because "I just want to" or "I'd love to use this new tech" or "I just think it's better." Sometimes the cost for these things is small, and sometimes it's high. (New tech is almost always very high.) So I sleep on PRs, I get input from others, and I try to be mindful of myself.
I'm concerned about the true purpose behind this PR and this effort. That concern would not exist if you weren't on such a hard-line position. Just say something to the effect of, "I think Hashie is a very expensive dependency, and I think the performance and API improvements are worth the cost of removing it." See, then you make sense, and you're demonstrating good programming. But what I see is: "HASHIE SHOULD NEVER BE USED IN PRODUCTION!!!", untested code, and arguments that more code is less code.
I mean, gosh... imagine counting code like you say. I need a simple CRUD app for managing a list of my cats. Just their names and breeds. Well, obviously, I can't use Rails, because... gosh, my one-model Rails application would have hundreds of thousands of lines of code! But nobody counts LOC this way, for many good reasons that I won't get into during the Merry Christmas season.
That's not the coverage I mean. If I'm testing a method like this
def do_something a b c end
and I test for
Without running the tests, I'd bet OmniStruct is covered well in the first sense. But in order to actually be tested is would need detailed tests at this level: https://github.com/intridea/hashie/blob/f42367e30e6a868ce117b143729b02d5f2c0a279/spec/hashie/mash_spec.rb
Anyway, I just wanted to offer my thoughts, which I did. I'm going to do more Christmas-y things now and stop bugging you. Merry Christmas!
Maybe you missed my statement earlier
This is conversation is neither appropriate or productive. You're attacking me in a very thinly veiled manner, you're arguing against a CLOSED PULL REQUEST, and generally taking up the time I could be spending contributing to the community. I don't need to be -d right now. I'm unsubscribing from the issue. Have a great end of the year
@darrencauthon You seem to just be playing devils advocate because @schneems is very clear about his position. @schneems has been over-the-top explicit why he holds his opinion. If you have any specific concerns to his points, I suggest you raise them instead. Your current comments are just vague theorizing about tradeoffs.
Frankly, @schneems has done more than enough to convince me of his stance. To see the forest through the trees here.
The raw PR numbers are +112 / -32. The largest class addition is 72 lines. We're hardly talking about massive unmaintainable classes here, let alone "large" and "complicated"
The crux of this argument is removing Hashie and replacing it with either a Hash, OpenStruct, or a well defined PORO. He is not proposing a C-extension or cryptic ruby code where one wrong move will easily invalidate the performance gain. And then boom, the performance of thousands of apps have been improved by ~5% for free (from their perspective).
I think that bears repeating: A change to remove a dependency and use idiomatic Ruby code improves the performance of thousands of apps by ~5%.
Without any concrete arguments, I don't know why this effort isn't met with more appreciation for @schneems's hard work investigation and then proposing a proof of concept. I know I am convinced.
For what little it's worth, as a builder of a several strategies and a huge advocate of OmniAuth for many years, I'm very supportive of removing Hashie for performance benefits sake. I'd be more than happy to rewrite my strategies to be able to pass on that kind of performance increase to any application using my libraries.
Also, huge thanks to schneems for putting in 16+ hours of effort on a very thoroughly investigated alternative and well thought out PR. Hopefully this leads to big gains in OmniAuth in it's next major release.