Fast Commits for Ext4 | Hacker News Reader

Fast Commits for Ext4

lwn.net

93 points_xrjp7 comments

Previous discussion (51 comments) https://news.ycombinator.com/item?id=25794659

Great development! I'm certainly all for this becoming the standard/default. Unfortunately for those relying on fsync() flushing the whole file system (which one shouldn't) syncfs() isn't a great replacement. Last time I tested that it didn't return errors on failure nor was it making certain all changes to the file system before syncfs() made it to disk.

Btrfs had this "fast fsync" for a while already btw. (out of necessity -- otherwise fsync would be even more unbearably slow on btrfs). This seemed to be difficult to implement (lots of corner cases). Every release there were "fast fsync" bug fixes such as e.g. https://lore.kernel.org/linux-btrfs/20200220132949.20571-1-f... https://lore.kernel.org/linux-btrfs/20200722112837.15516-1-f... . Let's hope the ext4 version of this has less initial issues.

I didn't realise fsync syncs the whole fs. This explains some write cache issues I've been having writing real-time data streams to sometimes slow SD cards. This change would solve my problems rather than create problems. But hey now I can fix it anyway

fsync is supposed to sync the entire filesystem, yes. On a SD card, I would expect it to do what it says it does.

However, on regular hard drives, filesystem devs and hardware vendors are all incentivized to cheat, since io benchmarks are big and obvious, while dropping data on the floor with power loss or failing hardware is subtle and hard to notice.

In practice, reliably syncing a file can be surprisingly hard: https://danluu.com/file-consistency/

Anyone have links to the best resources to master the core paradigms of file systems and stay up to date with future R&D?

Best books? Websites? Journals? People on Twitter to follow? GitHub projects to watch? YouTube videos?

I would recommend reading all the fantastic storage papers coming out of WISC under Remzi and Andrea Arpaci-Dusseau:

http://pages.cs.wisc.edu/~remzi/research/

And since you're looking for a GitHub project to watch, I'll go ahead and share TigerBeetle since I'm working on that :) and since we're implementing many of these new storage ideas (disentangling journal corruption from system crashes, protocol-aware recovery for consensus-based storage, hash-chaining to detect misdirected writes and disk corruption, and Direct I/O to recover from the kernel page cache going out of sync with the disk):

https://github.com/coilhq/tigerbeetle

And then finally since you also asked for a YouTube video :) here's a talk we did on Thursday for the Mojaloop Foundation, discussing many of these ideas around storage safety and also around optimizing I/O performance (io_uring, the cost of syscalls and context switches, cache line aligned structures, Direct I/O) and putting this all together into a replicated state machine with strict serializability:

https://www.youtube.com/watch?v=6D7B6v06mBo

If you read a couple of key papers, you'd be well introduced to the fundamentals of linux in-kernel filesystems.

File systems like ext4 are descended from this approach: https://docs.freebsd.org/44doc/smm/05.fastfs/paper.pdf

File systems like btrfs, f2fs, etc are descended from this approach (log structured file systems): https://web.stanford.edu/~ouster/cgi-bin/papers/lfs.pdf

For cloud based storage, the LSM paper (w/ leveldb), and HDFS are good foundational references imho.

Obviously there are a lot of details beyond those areas, but understanding these are fairly short but will give you a basis to get much further to more cutting edge R&D discussions.

Reply on news.ycombinator.com