SQLite: 35% Faster Than the Filesystem
sqlite.org
sqlite.org
Makes sense when you consider you're throwing out functionality and disregarding general purpose design.
If you use a fuse mapping to SQLite, mount that directory and access it, you'd probably be very similar performance (perhaps even slower) and storage use as you'd need to add additional columns in the table to track these attributes.
I have no doubt that you could create a custom tuned file system on a dedicated mount with attributes disabled, minimized file table and correct/optimized block size and get very near to this perf.
Let's not forget the simplicity of being able to use shell commands (like rsync) to browse and manipulate those files without running the application or an SQL client to debug.
Makes sense for developers to use SQLite for this use case though for an appliance-type application or for packaged static assets (this is already commonplace in game development - a cab file is essentially the same concept)
Related ongoing discussion, if someone cares to test this:
For example, Ceph uses RocksDB as their metadata DB (and it's recommend to put it) directly on a block device, with the WAL on yet another separate raw device
https://docs.ceph.com/en/latest/rados/configuration/bluestor...
mke2fs -t ext4 -b 1024 -N 100000 -O ^has_journal,^uninit_bg,^ext_attr,^huge_file,^64bit [/dev/sdx]
(smaller block size, 100,000 inode file table entries (tuned to the number of blobs), no journal, no checksumming, no extended file attributes, use smaller integer file offset IDs, 32 bit padded vs 64 bit)
Then mount it and run the same test.
You could go even further and tune fopen BUFSIZE to be no greater than 12,000 bytes. You can even create this mount on a file inside your existing mount... which is essentially akin to having an sqlite file without needing a client library to read/write to it.
Anyway - if the purpose is to speed up reads and save disk space on small blob files, there is little need to ditch the file system and it's many many upsides.
https://unixdigest.com/articles/battle-testing-php-fopen-sql...
Has been on HN lately..
Let's also note the 4x speed increase on windows 10, once again underlining just how slow windows filesystem calls are, when compared to direct access, and other (kernel, filesystem) combinations.
This is a great talk on the topic https://youtu.be/qbKGw8MQ0i8?si=rh6WJ3DV0jDZLddn
To pile onto NTFS, it's performance is so notoriously bad that there are developer teams working on Windows projects that configure their build farm to do cross builds from Linux to Windows just to avoid the performance penalty.
I argued for a Linux laptop, and the boss said, "OK, prove it. Here's two equivalent laptops, time it.".
Turns out there was zero difference, or negligible (Windows won), between compilation times. That has always annoyed me.
I think there was something seriously flawed in your test. If you Google for a minute, you find multiple posts on how moving the same builds to Linux led to performance improvements in the range of 40% drops in build times.
Some anecdotes even compare doing the same builds in Ubunto with NTFS to see double-digit gains.
NTFS is notoriously awful in scenarios involving reading/writing many small projects. This is the bottleneck in Windows builds. There is a myriad of benchmarks documenting this problem.
Nowadays there are plenty of cross-platform projects to serve as benchmarks. Checking this can be as easy as checking out a project, start a full rebuild, and check how long it takes.
I don't understand what point you're trying to make. Should we give equal credit to unsuccessful initiatives when their failure is due to screwing up a critical part of the project?
I mean, the successful attempts document what is possible. What do you expect to report when you couldn't even manage to get something working?
Assume bloggers blog mainly about content that contains a positive message. I'm asserting that people blog more readily about their success than their failures.
So when you look at the blog literature, your population is not N, it is M. You don't see the failures because they don't tell the tale.
Here's a talk on porting rustup to Windows: https://www.youtube.com/watch?v=qbKGw8MQ0i8
To begin with, it takes rustup 3m30s to install on Windows. After rejigging the code again and again away from its naive implementation which works fine on Linux, to perform sympathetically towards NTFS, it takes the same rustup 14s to install. That's quite a performance gain! But it needed a lot of changes to rustup, and to Windows itself.
In any case it wasn't much of a fun codebase. But I think a good lesson was, always test it, always measure. Not casting shade on either OS.
I believe you can lock a Linux box down tighter than a windows box, but then you're trading compile times for other costs.
In your particular corporate environment that might be the case, but not in this case, I had free run of a fresh install and no offensive AV there, and detuned to remove the crap.
Other posters have said certain optimizations (which I'm not sure would help, it was pure compilation, no large files that I'm aware of). Just saying, always good to keep an open mind.
If, however, your tests make any filesystem calls or fork a child process, there’s slim chances that Linux doesn’t absolutely trounce Windows.
To throw in a morsel of anecdata: same laptop with a dual boot runs the PhotoStructure core test suite (some 11,000 tests which have a pleasing melange of system and integration tests, not just unit tests) in 8m30s on Windows 11. Almost all the same tests complete in 3m30s on Linux.
Anecdotally though, git is unbearably slow under Windows and compiles make all filesystem operations lag while I have never seen such problems under Linux.
There are other old-school techniques which are far easier to implement and maintain, such as using RAM drives/partitions. Expensing 32GB of RAM is simpler than maintaining weird NTFS configurations.
Splitting your project into submodules/subpackages also helps amortize the impact of long build times. You can run multiple builds in parallel and then have a last build task to aggregate the all. Everyone can live with builds that take 5 minutes instead of 3.
At one point they planned to replace it with a database filesystem, but that was too complicated and abandoned. That was probably the end of replacement work on NTFS. https://en.wikipedia.org/wiki/WinFS
Sans the API piece, think of it like storing blobs in SQL Server, just like SharePoint does.
I was lucky enough to play around with beta 1. Not much you could do with it, though.
I’ve benchmarked deleting files (around ~65,000 small node_modules sort of files) and it takes 40 seconds through Explorer, 20 seconds with rd, and roughly a second inside WSL2 (cloned to the VM’s ext4 virtual hard drive).
The issue is Defender in sync mode/other AV/other file system filters.
DevDrive as noted by default uses an async scanning technique as well as ReFS. ReFS will suffer the exact same performance issues with Defender (or other AV/other file system filters) doing its thing when running in sync mode, which it does by default for ReFS-formatted drives in Windows Server.
https://gregoryszorc.com/blog/2021/04/06/surprisingly-slow/
https://news.ycombinator.com/item?id=26737521
> Except for CloseHandle(). These calls were often taking 1-10+ milliseconds to complete.
> While I didn't realize it at the time, the cause for this was/is Windows Defender. Windows Defender (and other anti-virus / scanning software) typically work on Windows by installing what's called a filesystem filter driver.
This doesn't take away from your point that _it is slow_, but the reasons are not due to the file system in use.
I've had folders take a full minute to open on an SSD.
It got to the point where I went to open the folder, it started loading. I needed the file quickly, so I searched for it online, found it, and opened it before windows finished loading that folder for me.
After exempting that folder from Windows Defender the folder loads instantly. For the life of me I cannot understand why Defender blocks Explorer.
I suppose if you wanted to find out, you could use dtrace/ETW.
Explorer has other things going on, though, including other apps that hook into it (shell extensions, like Adobe Reader, TortiseGit/SVN, and so on) which can certainly cause performance issues.
Just one example: File icons or thumbnails can be dynamically generated by shell extensions based on the file contents. A maliciously crafted file could potentially exploit a vulnerability in such a shell extension.
Windows has an extensible model. It's a different approach from most (all?) other OSes. It offers a different set of features.
Sure, AV could perhaps be done in a different manner that would be more effective/faster, I can't comment on that as I lack the insight required -- only MSFTies that work on kernel code could respond in any authoritative way.
But I'm not particularly knowledgeable either on this topic, just a (forced) consumer of the operating system with the occasional reading on the side
FAT[32] does implement minifilters.
https://learn.microsoft.com/en-us/windows-hardware/drivers/i...
https://www.osr.com/nt-insider/2019-issue1/the-state-of-wind...
I guess my mental model on this was just lacking/simply wrong.
If you want to know more, grab the book Windows Internals.
The kicker... Linux was running inside a VirtualBox VM inside the very same Windows host.
This could also be some variance in the `javac` command between OS's, granted.
https://devblogs.microsoft.com/visualstudio/devdrive/
But the primary improvement comes from async A/V, not the file system.
ReFS offers other improvements such as CoW and instant file initialization which can benefit developers. From this standpoint, it is the correct choice over NTFS.
https://devblogs.microsoft.com/engineering-at-microsoft/dev-...
"Are you done yet?" vs "Get back to me when you're ready, gonna go do something else". Aka blocking i/o vs nonblocking i/o.
(I always wonder why Windows NT doesn't get more high performance implementations for networking et. al. with this feature; licensing? trade off in perf elsewhere? can't tweak kernel params to your heart's desire?)
io_uring implemented an additional feature of a ring buffer, but Microsoft has followed this feature; I think it was first introduced in a later build of Windows 10 and should be in 11 & 2022.
Microsoft has a graphical example of IOCP:
https://learn.microsoft.com/en-us/windows/win32/fileio/synch...
More info on a comparison of ring implementations in Windows/Linux:
https://windows-internals.com/ioring-vs-io_uring-a-compariso...
https://forum.lazarus.freepascal.org/index.php/topic,66281.m...
One could flip it around and store logs in a multimedia container, but then you won't have nice indices like with sqlite, just the one big time index
Seriously though, I think this is a great idea, and would be interested in how easy it is to write sqlite output adaptors for the various logging libraries out there.
And they won’t be wrong.
sqlite3 logs.db "select log from logs" | grep whatever
https://www.sqlite.org/c3ref/update_hook.html
On a PDP-11, run this program via telnet, rsh, or rexec.
If you're more ambitious, porting SQLite to 2.11BSD would be a fun exercise.
Maybe a user defined function bound to an INSERT trigger. But I believe that functions are also connection specific which is fine for the tail tool, but what happens when the user defined function connection goes away.
I wish Splunk and friends would have an interface like that. Sure it does basic grep, and it is a much more powerful language, but sometimes you just needed some command line magic to find what you wanted.
https://git.sr.ht/~martijnbraam/logbookd
Although I'm not sure it uses WAL2 mode, but that should be a trivial change.
ClikcHouse works really well when you need to store and analyze large logs but compare to SQLite it would require to maintain a server(s). There is DuckDB which is embedded like SQLite and it could be a better than SQLite fit for logs but I have no experience with DuckDB.
I am not sure which one would be better for logs, I would need to play around with it. But i am not sure if SQLite wouldn’t be a better fit.
I was debating what storage layer to use and decided to try SQLite because of its speed claims — essentially a single table where each row is a MIDI event from the piano (note on, note off, control pedal, velocity, timestamp). No transactions, just raw inserts on every possible event. It so far has worked beautifully: it’s performant AND I can do fun analysis later on, e.g. to see what keys I hit more than others or what my average note velocity is.
It’s also worth taking into consideration damper pedal velocity changes. When you go from “off” (velocity 0) to fully “on” and depressed (velocity 127), a lot of intermediate values will get fired off at high frequency.
Ultimately though you are right; it’s not enough frequency of information to overload SQLite (or a file system), probably by several orders of magnitude.
I wonder how io_uring compares.
I've used SQLite blob fields for storing files extensively.
Note that there is a 2GB blob maximum: https://www.sqlite.org/limits.html
To read/write blobs, you have to serialize/deserialize your objects to bytes. This process is not only tedious, but also varies for different objects and it's not a first-class citizen in other tools, so serialization kept breaking as my dependencies upgraded.
As my app matured, I found that I often wanted hierarchical folder-like functionality. Rather than recreating this mess in db relationships, it was easier to store the path and other folder-level metadata in sqlite so that I could work with it in Python. E.g. `os.listdir(my_folder)`.
Also, if you want to interact with other systems/services, then you need files. sqlite can't be read over NFS (e.g. AWS EFS) and by design it has no server for requests. so i found myself caching files to disk for export/import.
SQLite has some settings for handling parallel requests from multiple services, but when I experimented with them I always wound up with a locked db due to competing requests.
For one reason or another, you will end up with hybrid (blob/file) ways of persisting data.
https://medium.com/@rishabhdevmanu/from-trees-to-tables-stor...
This makes total sense and it is also "frowned upon" by people who take a too purist view of databases
(Until it comes a time to backup, or extract files, or grow a hard drive etc and then you figure out how you shot yourself in the foot)
Also note that SQLite does have an incremental blob I/O API (sqlite3_blob_xxx), so unlike most other RDBMS there is no need to read/write blobs as a contiguous piece of memory - handling large blobs is more reasonable than in those. Though the blob API is still separate from normal querying.
MS SQL Server: READTEXT, WRITETEXT, substring, UPDATE.WRITE
Oracle: DBMS_LOB.READ, DBMS_LOB.WRITE
PG: Large Objects
Most of my experience is with SQL server and it can stream large objects incrementally through a web app to browser without loading the whole thing into memory at 100's Mbytes/sec on normal hardware.
image and text with READTEXT and WRITETEXT are deprecated but still work fine, varbinary(max) and substring UPDATE.WRITE are the modern equivalent and use the same implementations underneath.
Filestream allows larger than 2 gigs, stores the blob data in the filesystem but otherwise is accessed like the other blobs along with some special capability like getting a SMB file pointer from the client for direct access. It is also backed up and replicated like normal, definitely not just like storing a path in the DB.
Filestream performs worse the in db for blobs under about 100kb I believe where it would be recommended to keep them in db for max perf.
I have used MSSQL blobs long before filestream existed and it works well except for the 2 gig limit and once the db gets large backup and log management get more unwieldy than if you just stored them outside the db but it does keep them transactional consistent, which filestream also does.
I'm confused what you mean by this. Files also only contain bytes, so that serialization/deserialization has to happen anyway?
For example, `pd.read_parquet` accepts "file-like objects" as its first argument: https://pandas.pydata.org/docs/reference/api/pandas.read_par...
However, this is not the case for fringe tools
Could use a named pipe.
I’m reminded of what I often do at the shell with psub in fish. psub -f creates and returns the path to a fifo/named pipe in $TMPDIR and writes stdin to that; you’ve got a path but aren’t writing to the filesystem.
e.g. you want to feed some output to something that takes file paths as arguments. We want to compare cmd1 | grep foo and cmd2 | grep foo. We pipe each to psub in command substitutions:
diff -u $(cmd1 | grep foo | psub -f) $(cmd2 | grep foo | psub -f)
diff -u /tmp/fish0K5fd.psub /tmp/fish0hE1c.psub
ksh and bash too have this as <(…) and >(…) under Process Substitution.
An example from ksh(1) man page:
paste <(cut -f1 file1) <(cut -f3 file2) | tee >(process1) >(process2)diff <(cmd1 | grep foo) <(cmd2 | grep foo)
This is a silly argument, there's no reason to recreate the full hierarchy. If you have something like this:
CREATE TABLE files (path TEXT UNIQUE COLLATE NOCASE);
SELECT path FROM files WHERE path LIKE "./some/path/%";
The reason it's fast is because the table has a UNIQUE index, and LIKE uses it if you turn off case-sensitivity. No need to faff about with hierarchies.
EDIT: btw, I am using SQLite for this purpose in a production application, couldn't be happier with it.
i'm sure your suggestion exhibits creative step-one thinking, but i've described several reasons why it doesn't make sense to recreate a filesystem in sqlite, the combination of which should make it clear why doing so is naive
(1) Slim table "items"
- id / parent_id / kind (0/1 file folder) integer
- name text
- Maybe metadata.
(2) Separate table "content"
- id integer
- data blob
There you have file-system-like structure and fast access times (don't mix content in the first table)
Or, if you wish for deduplication or compression, add item_content (3)
In the process of prototyping some "remote" collaborating file systems, I always wonder whether it is a good idea maintaining a flat map from path concatenated with "/" like an S3 to the file content, in term of efficiency or elegancy.
It surprised me that almost all required PostgreSQL, and most of those that didn't opted for something otherwise complex such as Mongo or MySQL.
SQLite, with no dependencies, would have simplified the process no end.
All fs/drive access is managed by the OS. No DB systems have raw access to sectors or direct raw access to files.
Having a database file on the disc, offers a "cluster" of successive blocks on the hard drive (if it's not fragmented), resulting in relatively short moving distances of the drive head to seek the necessary sectors. There will still be the same sectors occupied, even after vast insert/write/del operations. Absolutely no change of DB file's position on hard drive. It's not a problem with SSDs, though.
So, the following apply:
client -> DB -> OS -> Filesystem
I think, you already can see the DB part is an extra layer. So, if one wouldn't have this, it would be "faster" in terms of execution time. Always.
If it's slower, then you use the not-optimal settings for your use case/filesystem.
My father did this once. He took H2 and made it even more faster :) incredible fast on Windows in direct comparison of H2/h2-modificated with same data.
So having a DBMS is convenient and made in decisions to serve certain domains and their problems. Having it is convenient, but that doesn't mean it's the most optimized way of doing it.
Oracle can use raw disks without having a filesystem on them, though it's more common to use ASM (Automatic Storage Management) which is Oracle's alternative to raw or OS managed disks.
https://docs.oracle.com/en/database/oracle/oracle-database/1...
Indeed. Offers 10-12 percent performance writing direct on a block device. But, this block device is still attached to a running system :)
* https://news.ycombinator.com/item?id=14550060 7 years ago
* https://news.ycombinator.com/item?id=20729930 5 years ago
* https://news.ycombinator.com/item?id=27137834 3 years ago
* https://news.ycombinator.com/item?id=27897427 3 years ago
* https://news.ycombinator.com/item?id=34387407 2 years ago
In my opinion, SQLite can be faster in big reads and writes too, but the team didn't optimise it as much (like loading the whole content into memory) as maybe it was not the main use of the project. My hope is that we will see even faster speeds in the future.
[1] https://pack.ac [2] https://forum.lazarus.freepascal.org/index.php/topic,66281.m...
What is SQLite not doing that filesystems are?
Access control is a big cost. Some AV systems (like everyone's favourite Crowdstrike) also hook every open/close.
I am not sure if this is done anymore, because the performance gains were modest compared to the hassle of a custom formatted partition.
It's important. But understandable.
It seems to me that all the extra indirection from using FUSE would lead to more than a 35% performance hit.
Statically linking an sqlite into a kernel module and providing it with filesystem access seems like something non trivial to me.
Putting a relational database in the OS kernel is an interesting violation of standard layering. Obviously has the potential to unleash a lot of issues, but also could possibly enable novel features.
> The performance difference arises (we believe) because when working from an SQLite database, the open() and close() system calls are invoked only once, whereas open() and close() are invoked once for each blob when using blobs stored in individual files.
I wish I could find that comment, because my explanation doesn't do it justice. Very interesting idea, someone's probably going to explain how it's already been tried in some old IBM database a long time ago and failed due to whatever reason.
I still think it should be tried with newer technologies though, sounds like a very interesting idea.
The original article effectively argues the opposite: if your use case matches a database, then that will be way faster. Because the filesystem is both fully general, multi-process and multi-user, it's got to be pessimistic about its concurrency.
This is why e.g. games distribute their assets as massive blobs which are effectively filesystems - better, more predictable seek performance. Starting from the Doom WAD onwards.
For an example of databases that use the file system, both the mbox and maildir systems for email probably count?
They basically mmap’ed the database file and argued that OS cache should do its job. Which makes sense but I guess it did not perform as well as any fune tuned caching mechanism.
It's now "somewhat possible" in that modern filesystem are overall mostly less broken about handling large number of small (or at least moderately small) files than they used to be.
But databases are still far more optimized for handling small pieces of data in the ways we want to handle data we put into databases, which typically also includes a need to index etc.
It ships with an example VFS which shows you how to do this: https://www.sqlite.org/src/doc/trunk/src/test_onefile.c
Using SQLite let's you tailor your data access patterns in a much more rigorous way and side step the POSIX tarpit.
I guess querying by PK has some similarities but it is not as unstructured and random as a seek.
Also side effects such as sparse files do not mean much from a database interface standpoint.
https://www.sqlite.org/c3ref/blob_open.html
https://www.sqlite.org/c3ref/blob_read.html
I've not seen performance numbers for those. Could make for an interesting micro-benchmark.
What do you expect the value proposition of something loosely described as a sqlitefs to be?
One of the main selling points of SQLite is that you can statically link it into a binary and no one needs to maintain anything between the OS and the client application. I'm talking about things like versioning.
What benefit would there be to replace a library with a full blown file system?
https://github.com/narumatt/sqlitefs
And it seems quite interesting:
"sqlite-fs allows Linux and MacOS to mount a sqlite database file as a normal filesystem."
"If a database file name isn't specified, sqlite-fs use in-memory-db instead of a file. All data will be deleted when the filesystem is closed."
Because with OpenBSD introduction of pinning all syscalls to libc, doesn’t this block SQLite from making syscall direct.
not a native speaker, what does it mean?
(Sometimes the phrase is casually used to just mean "a lot", but here I think they mean 10x).
Which is a bit d'oh, since being faster for some things is one of the main motivations for a database in the first place.