1.4.1 and more

Mon 25 February 2013 by Monnand

Today, I have released uniqush-push 1.4.1, whose release note is available now along its binary files. It is highly recommended to update to this version, since there will be some non-deterministic bugs in old versions. In this blog, I will first go through the changes and explain some of them in details.

The changes in this version contains some improvements and bug fixes:

  • improvement On (un)subscription, the log message now contains service and subscriber information. Fixed issue 2.
  • improvement On adding an APNS push service provider, uniqush-push will report error immediately if the given certificate or key files do not exist. Fixed issue 6.
  • bugfix Removed all code related memory pool. There is no dependency on memory pool any more. Fixed issue 7.
  • improvement Rewrote the code in main package. It is more clean, simple and less redundant. Fixed issue 5.

I will explain them one by one below.

First two improvements

These two changes are not quite important but useful in its own right.

Apparently, adding more information into the log will help users and myself to identify the potential bug, which is generally a good thing as long as we do not leak sensitive information.

A system would be less error-prone if it reports error as soon as possible. Before this version, users can add an APNS push service provider even by providing a wrong path to the certificate and/or key files. The user can only know this error if they are actually using the malformed APNS provider to push something to a client. This is too late, and it should be better to report the error in advance.

No more memory pool

This is a hard decision to make.

I came up with the memory pool idea because of the famouse blog post. In the post, Russ Cox was able to optimize a Go program 15x faster than the original version, making it even faster than an equivalent C++ program. That post is amazing and I really like it.

In the blog post, Russ Cox said:

Having a garbage-collected language doesn't mean you can ignore memory allocation issues. In this case, a simple solution is to introduce a cache so that each call to FindLoops reuses the previous call's storage when possible.

Given the C/C++ background I had, the sentences above, to me, were like saying: use a memory pool under whatever condition if you want to have a fast program.

Then I made a memory pool. That is the begining of the nightmare. I received several bug reports on the mailing list saying that they observed some very weird behavior. For example, some client received a message which should be sent to another one. I tried to re-produce their bugs bug failed.

Finally, I reallized that the push service provider and delivery point structures were not cleared before putting back to the memory pool. This reminded me about the weird bugs on the mailing list.

I had two options: clear all data before putting them back to the pool; or drop the whole idea of memory pool.

Well. I do like a fast program, but I love simplicity as well. Managing my own memory using a garbage-collected language became a very awkward behavior to me. It is true that Go's garbage collector is not perfect right now. Even if it is good enough, it may need a good care of memory when writing programs anyway. But it is not my current concern. I want a simple program which can work. In fact, I even wrote a version with a correct behavior using memory pool. But I decided to not release it because of potential bugs caused by complexity.

Refactory

OK. this is a strong word for developers. Yes, I rewrote the whole main package.

I have to say that developing uniqush is really a learning process to me. I developed quite a few programs, some of them are even widely used. However, I haven't taken care of a project before uniqush. In uniqush, I have to write correct code while making sure it can envolve in a healthy way. This includes correcting my own mistakes and bad decisions in the code.

Refactorying the whole main package is correcting my bad design decision. I made a wrong choice of using channels and goroutines in Go making the main package cumbersome and hard to maintain. After this refactoring, the whole main package contains only four files and reduced around a thousand of lines of code. (You can see this fact from the code frequency.)

Besides, since I am going to support feedback service in APNS, such refactorying in inevitable. The service requires a new interface for push service providers and old main package can hardly support the new interface.

What about... uniqush-conn?

Sorry! I didn't expect that some essential parts are missing in Go and I have to write them by my own. More specifically, I have already written a Diffie-Hellman Key Exchange library in Go, and I am going to implement RSASSA-PSS signature algorithm later. Even though, I still want to use Go on server side, simply because it is too much fun.

With that being said, uniqush-conn will be delayed, again.


Comments