Matrix's reference homeserver, Synapse, is written in Python and uses the Twisted networking framework to power its bitslinging across the Internet. The Python version used has been strictly Python 2.7, the last supported version of Python 2, but as of this week that changes! Since Twisted and our other upstream dependencies now support the newest version of Python, Python 3, we are now able to finish the jump and port Synapse to use it by default. The port has been done in a backwards compatible way, written in a subset of Python that is usable in both Python 2 and Python 3, meaning your existing Synapse installs still work on Python 2, while preparing us for a Python 3 future.
The Python 3 port has benefits other than just preparing for the End of Life of Python 2.7. Successive versions of Python 3 have improved the standard library, provided newer and clearer syntax for asynchronous code, added opt-in static typing to reduce bugs, and contained incremental performance and memory management improvements. These features, once Synapse stops supporting Python 2, can then be fully utilised to make Synapse's codebase clearer and more performant. One bonus that we get immediately, though, is Python 3's memory compaction of Unicode strings. Rather than storing as UCS-2/UTF-16 or UCS-4/UTF-32, it will instead store it in the smallest possible representation giving a 50%-75% memory improvement for strings only containing Latin-1 characters, such as nearly all dictionary keys, hashes, IDs, and a large proportion of messages being processed from English speaking countries. Non-English text will also see a memory improvement, as it can be commonly stored in only two bytes instead of the four in a UCS-4 “wide” Python 2 build.
Editor's note: If you were wondering how this fits in with Dendrite (the next-gen golang homeserver): our plan is to use Synapse as the reference homeserver for all the current work going on with landing a 1.0 release of the Matrix spec: it makes no sense to try to iterate and converge on 1.0 on both Synapse and Dendrite in parallel. In order to prove that the 1.0 spec is indeed fit for purpose we then also need Synapse to exit beta and hit a 1.0 too, hence the investment to get it there. It's worth noting that over the last year we've been plugging away solidly improving Synapse in general (especially given the increasing number of high-profile deployments out there), so we're committed to getting Synapse to a formal production grade release and supporting it in the long term. Meanwhile, Dendrite development is still progressing - currently acting as a place to experiment with more radical blue-sky architectural changes, especially in low-footprint or even clientside homeservers. We expect it to catch up with Synapse once 1.0 is out the door; and meanwhile Synapse is increasingly benefiting from performance work inspired by Dendrite.See 10/15, ~20:00 for the Python 3 migration. This is on some of the Synchrotrons on matrix.org.
See ~11/8 for the Python 3 migration. This is on the Synapse master on matrix.org.
We have also noticed some better CPU utilisation:
See 10/15, 4:00.The CPU utilisation has gone down on synchrotron 1 after the Python 3 migration, but not as dramatically as the federation reader. Synchrotron 3 was migrated a few days later.
As some extra data-points, my personal HS consumes about 300MB now at initial start, and grows to approximately 800MB -- under Python 2 the growth would be near-immediate to roughly 1.4GB.
Happy Matrixing,
Amber Brown (hawkowl)