MongooseIM 3.0.0 - Application turbocharger
By Piotr Nosek
MongooseIM 3.0.0 is out and with it come many improvements to our global messaging solution! Over the years we have proven that MongooseIM is the way to go when building a scalable, secure messaging system that never fails. With new features and fixes, our battle tested, highly customisable platform provides an enterprise friendly toolbox everyone can use. Whether you’re an XMPP expert or an entrepreneur looking to bring to life your idea for a community building app, MongooseIM platform helps you build a product tailored to your needs, that will easily grow to match your ambition. Find out what goodies we’ve managed to pack up for you this time and see how we can aid your users’ experience with our truly instant messaging platform.
What is so great about 3.0.0?!
As a team we’ve switched into a faster gear and our latest release is a reflection of that. MongooseIM 3.0.0 is an enterprise ready, stable solution that now works faster then ever; delivering a smooth messaging experience to your users. It features important upgrades that will allow your servers to process even more messages per minute and save memory space for additional user sessions. Everything thanks to a couple of improvements, the new XML parser being the prominent highlight.
Efficiency is not the only reason why we’re so proud of this release. It also features prototype Inbox implementation. It’s an essential extension for virtually every chat application. Our rich experience in this topic allowed us to design a solution that should match most use cases already and will be expended even further!
As usual, there are also other improvements we’d like to share with you. You may find the full list in our changelog but we’ve picked five of them to describe, as we feel you should learn more about them.
Achieve more with the same hardware
Thanks to several important changes and improvements, MongooseIM is now able to process information faster and consume less resources. It means your servers may handle more users and traffic with the MongooseIM upgrade alone.
Depending on your specific application, 25–400% better performance may be expected. Actually, the richer and more complex trafic, the better results you’ll get, compared to previous MongooseIM versions!
Three aspects of MongooseIM have been modified in order to achieve this:
- All messages from users are interpreted in a completely new way
- More users can connect to a single server per second
- All user sessions store as little information as possible when they are idle
We’ve implemented Inbox features in the past for various projects and the time has come to pour the best ideas and experiences into an extension open for everyone!
A few words of explanation for those not familiar with the Inbox feature. It is the view in a chat application that you see every time you open it. It is a list of all conversations, with excerpts of last messages and unread messages count. Simple as that!
Unfortunately, as there is no official Inbox specification in XMPP yet, we’ve come up with custom protocol (thank you XMPP for your extensiveness!) for this purpose. We’re going to submit it as a XEP (XMPP Extension Protocol) for a review by community but in the meantime you can enjoy its simplicity and intuitiveness. All you need to do is to enable mod_inbox and implement a few simple IQ stanzas in your client application.
Please keep in mind though, this extension is still in experimental stage and will be marked as stable in one of our future releases.
Under the hood
We’ve also added some lower level changes that are going to be useful to developers, CTOs and devops.
Performance bundle: Acceptor pool, session hibernation, new XML parser
We’ve been using expat parser since the very beginning. Recently, our brave C++ warrior in the MongooseIM team thought of replacing it with an alternative — RapidXML. Everything indicated that it might consume less resources if properly used.
And he did use it in an excellent way. The whole C code was rewritten into C++ with the new library integrated. Since RapidXML requires only minimal per-user state (which actually is kept in Erlang terms), much fewer allocations are required and the code is simply cleaner.
What about the performance itself? For smaller stanzas the difference is not drastic, however noticeable. The test involved 100,000 users sending standard, small messages.
Please take a look at CPU usage graphs first:
Very similar but new version is a bit better in overall.
Regarding the memory consumption:
Now that’s already impressive. 3.0.0 uses ~2GB of RAM less which is over 25% improvement!
Wait, wait. If that was “impressive” then how do we call this? 1000 users exchanging large, ~36kB messages. CPU goes first again.
Well, it’s only 4 times better. :) In terms of memory usage the drop is less significant but still observable.
To wrap it up: all applications will surely benefit from a new parser but it’s especially important for those which process complicated, nested, rich stanzas.
Obviously, rewriting thousands of lines of code in C++ isn’t always the only method of improving performance. Sometimes it’s about small ideas, such as using a pool of acceptors or eager hibernation.
The former involves using more than one process to accept incoming connections from clients. Despite being a fairly cheap operation (it’s only about accepting a connection and creating a client process), this single process became a bottleneck in some applications. In 3.0, 100 acceptors are created by default.
The latter is bit more complicated. For people less familiar with Erlang this might sound similar to Java’s Hibernate framework. Actually, Erlang’s hibernation is not related to database mappings. When a process gets hibernated, its memory is garbage collected and it is removed from a scheduler queue (a bit of a simplification, but you get the idea). We already had such a mechanism in place but the hardcoded timeout for hibernation was 60s. When you think about it, client’s process is idle most of the time, at least in a computers’ world where a single CPU cycle takes less than 1ns. Load tests have proved that hibernating immediately after processing a stanza leads to lower memory usage at minimal cost to CPU time.
The first two graphs show CPU usage in a presence-based test. Most of the time ~6.5k users were connected and sending a presence update every 20 seconds to a roster of 8 friends.
Fairly similar I’d say. The second graphs shows memory usage decrease in the same test.
Now, that looks definitely better, right? However, if it turns out that frequent hibernations impact your server’s performance, you may easily tune the timeout in the configuration file (look for the hibernate_after option).
Improved ODBC support
Before 3.0, MongooseIM was using the odbc application from OTP to execute queries via ODBC. Unfortunately, this library is not maintained actively enough to match our requirements. It especially applies to SQL types support. Lucky for us, there is a community-developed repository named eodbc. With its help, MongooseIM’s compatibility with e.g. MSSQL improved significantly. What is more, in order to ensure it, we’ve begun testing ODBC connection with MSSQL on Travis!
A byproduct of this refactor is a completely new escaping API in our RDBMS layer. It’s more intuitive now and much less error prone. Now it’s virtually impossible to use unescaped value in a query and the escaping is always done appropriately for a chosen RDBMS.
Farewell, Message Archive Management v0.2!
Getting rid of MAM 0.2 support means several things, such as easier code maintenance. For example, there were completely separate functions and tests present to handle 0.2 stanzas.
Another important difference is a lack of “archived” element. Please configure MongooseIM to inject “stanza-id” into messages instead.
Despite its importance (MAM 0.2 was the first version supported by MongooseIM), it has become obsolete over time. If your application still uses MAM 0.2, we highly recommend you update your XMPP library and the code using it. Storage backend wise, newer versions are backwards compatible.
Please feel free to read the detailed changelog, where you can find a full list of source code changes and useful links.
After introducing big, important changes over past few releases, we’re going to take our time to polish what we already have.
Above all else, the Inbox feature is going to be expanded. We’re planning to support more backends and introduce new functions (e.g. sorting by timestamp). Also, we’d like to propose a proto XEP to the XMPP community, so the conversations list we’ve design may become an official standard common to clients and other servers. After all, MongooseIM is not our only priority, as we care for the state of XMPP as a communication protocol as well!
We’re slowly heading towards a configuration file revolution. It’s highlight will be a new config format, that will be friendlier for everyone not familiar with Erlang syntax. Currently we’re considering YAML, TOML, Cuttlefish and Conform. What is more, the configuration will undergo a major cleanup and flexibility improvements.
The third item I’d like to share with you applies to everyone working with MongooseIM code. One of the main structures in our server, Mongoose Accumulator (or mongoose_acc) is going to be significantly refactored. Our aim is to make it more intuitive and organised. Its content will be richer, with clear contract and scope. We’re redesigning it as you read these words. :)
Test our work on MongooseIM 3.0 and share your feedback
- Help us improve the MongooseIM platform:
- Star our repo: esl/MongooseIM
- Report issues: esl/MongooseIM/issues
- Share your thoughts via Twitter: twitter.com/MongooseIM
- Download Docker image with new release
- Sign up to our dedicated mailing list to stay up to date about MongooseIM, messaging innovations and industry news.
- Check out our MongooseIM product page for more information on the MongooseIM platform.
Originally published at www.erlang-solutions.com.