Show HN: Syq – copy files between machines fast (better than rsync)

Posted by greaber 2 days ago

I got frustrated with the slowness of rsync and made an alternative that works faster by using multiple parallel connections, direct encrypted TCP when available, and other optimizations. I also added cool features like the ability to maintain a persistent ssh connection to the server for fast one-offs, the ability to download to your laptop while working in an ssh shell on a server, and the ability to do direct remote-remote transfers without forwarding your ssh agent (by using restricted ssh keys on the receiver that will only execute a specific request signed with the key on your laptop). It is also designed to be more robustly/flexibly scriptable than alternatives.

Comments

Comment by hxseven 2 days ago

I thought I'd share a more positive comment. I work on a tool in a similar area (moving stuff around) but on a higher level. I currently use rsync or rclone for the file moving parts. So to me your tool looks interesting and promising, thanks for sharing it :)

Some suggestions for improvement for your website/Github:

- I wanted to understand how it differs from rsync. A comparison page would be helpful.

- A comparison to rclone would also be interesting, as it also uses multiple parallel connections.

- As it target more advanced users I think it would be helpful to provide more advanced technical insights on how your transfers exactly work. Also how are edge cases handled, what tests are done? Moving and copying files is critical, and you need to inspire trust in your project among users.

Comment by greaber 1 day ago

Thanks! Yeah, I was a little surprised (maybe I shouldn't have been) at how negative some of the comments were. I will think about what additional explanations I can provide that will help people. Regarding rclone, I have also spent a bunch of time looking at faster transfers to and from S3-compatible storage, but that problem is somewhat different, and I didn't look in depth at how rclone is implemented. FWIW, in my experience, s5cmd is generally a bit faster than rclone even if you tune rclone options. (But rclone is more flexible.) My guess is that s5cmd is already close to the performance ceiling, but I don't really know.

As for edge cases, I will also try to document that more. In general, when I'm not sure what semantics to go for, I try to either copy what rsync does or do something safer. For instance, one thing I am looking at now is the best way to handle cases where a source path cannot be represented on the target filesystem or where two source paths would collide (e.g. due to unicode normalization or case insensitivity). Tentatively, my inclination is to try to fail before copying any files when possible. Currently, syq doesn't do this (but neither does rsync).

Comment by rdbl27 1 day ago

Don't despair; new tool rollouts are always a hard sell -- especially when replacing a beloved classic.

For me, the killer feature would be "ease of use."

Sure, I _can_ make rsync do any number of rarely-used optimizations, if I feel like studying the manpage for half an hour and figuring out how to fit it to my exact use case.

If your tool has the same features but is automatically adaptive -- I'd use it.

Copying thousands of tiny files in deeply nested subdirs? Just works. Copying a huge file that's already encrypted? Just works. Copying a mix? Just works.

No special flags to set, zero config. It just works, optimally, every time.

"Reducing the user's cognitive load" is the killer feature here.

Comment by unsnap_biceps 2 days ago

> I also added cool features like the ability to maintain a persistent ssh connection to the server for fast one-offs

rsync already supports this via OpenSSH's ControlMaster directive. Bonus, it speeds up every connection to that server rather then a single tool's.

> the ability to do direct remote-remote transfers without forwarding your ssh agent (by using restricted ssh keys on the receiver that will only execute a specific request signed with the key on your laptop).

You can control your agent forwarding in your local ssh client configuration. And with modern ProxyJump, you don't need to forward your agent at all regardless of how many bastion hops are between you and your final target.

Comment by greaber 2 days ago

Yeah, ControlMaster is cool and was an inspiration, but since the syq client is always talking to the same receiver on the server, we can save a few round trips that ControlMaster still needs, so syq persist feels faster. This architecture is also what enables reverse mode, where you can download to your laptop while working on the server.

ProxyJump doesn't help when you want to copy files from server A to server B without giving server A an agent that can do arbitrary things on server B. There is really no alternative to using a restricted authorized key on server B for this scenario.

Comment by AceJohnny2 2 days ago

Thank you for showing this. I'll echo what some others say: if you're claiming to be faster than rsync, you should demonstrate why front-and-center. There have been decades of documentation on how rsync optimizes transfers, so it's a lot to live up to.

As an aside, It's perhaps an indictment of our networking landscape that multiple parallel connections between 2 specific machines would accelerate a transfer. I would've expected a single TCP connection to be able to saturate a line. Or perhaps the parallel connections seeks to amortize the per-file setup overhead?

Comment by greaber 1 day ago

I entirely agree that multiple connections shouldn't actually be necessary for speed, but they help in multiple situations (long-distance transfers, same DC between servers, NFS), and it's not just about amortizing per-file setup overhead.

There is some info on the optimizations in the docs, but I agree that a more complete technical explanation of all the things syq does could be useful. I will work on one. On the other hand, I also tried hard to make it just go fast without needing the user to understand why it is fast or tune anything. For instance, the number of connections is auto-tuned by default.

Comment by AceJohnny2 1 day ago

> There is some info on the optimizations in the docs, but I agree that a more complete technical explanation of all the things syq does could be useful

First, thank you for taking the feedback :)

To be clear, I'm focusing on the first thing you should be telling users. It seems that, as a CLI program, the target audience of this tool is fairly technical users who are likely already familiar with SSH, rsync, and everything. As demonstrated in this thread, your first communication task is to convince them why it's worth their time to use your tool, considering they're probably satisfied with the pre-existing ones (or a replacement would already exist!).

My point is, explaining why it's better/faster than rsync shouldn't be buried in the docs. It should be a paragraph on the front page.

Hope this helps, and keep it up. HN is a rough audience to go public on ^^;

Comment by fmajid 1 day ago

Multiple connections totally help. At a previous job, we drove rsync with GNU parallel, then did a final catch-all pass with rsync, and that divided transfer time for about 600GB by a factor of 4.

Comment by ranger_danger 1 day ago

> I would've expected a single TCP connection to be able to saturate a line

There's several factors at play that make this (usually) not the case.

Besides physical latency (which includes those added by any VPNs/tunnels/etc., some of which may be internal to an ISP along the route and outside of your control), there's other things like the TCP window sizes / window scaling option[1] that can affect single stream performance, and those type of parameters can differ by OS/interface type on both ends.

Also for SSH specifically, it has its own fixed buffer size that also limits throughput unless you're using the HPN-SSH fork[2].

[1] https://en.wikipedia.org/wiki/TCP_window_scale_option

[2] https://www.psc.edu/hpn-ssh-home/introduction/

Comment by greaber 1 day ago

Also, in syq, a connection is the unit of independent transfer work. Probably it would be possible to get higher bandwidth per TCP connection by some kind of multiplexing, but that architecture would be more complex and probably not have many benefits.

Comment by ranger_danger 1 day ago

There's also MPTCP (multipath TCP) which is even supported on mobile devices these days. So for example if your device has both a wifi and a cellular connection, and both ends of the stream support MPTCP, you can increase throughput/failover by combining multiple interfaces.

Comment by cynicalsecurity 2 days ago

I'm eager to replace battle-tested rsync with a new half-baked tool that no one uses for a tiny speed improvement. I'm also going to enable automatic updates of that tool too.

Comment by greaber 2 days ago

I'm seeing 5x speedups on my real workloads, and you can see some synthetic benchmarks here or run your own https://greaber.github.io/syq-bench/

Comment by throw1234567891 1 day ago

what's a "synthetic benchmark", claude came up with those numbers by looking at the code?

Comment by greaber 1 day ago

They are benchmarks on real hardware, and I tried to provide enough details so that anyone can reproduce them (but some of the benchmarks are on servers I have that you can't order on public cloud -- these ones are clearly marked). "Synthetic" means that the data is synthetic, not an actual production workload. For instance, some of the benchmarks use files filled with random bytes.

Comment by throw1234567891 1 day ago

So real benchmarks with synthetic data.

Comment by beepbooptheory 1 day ago

Do non-randomly ordered bytes produce different results??

Comment by greaber 1 day ago

Syq has compression turned on by default. Also, if you are updating a file that already exists on the other end, syq will only send blocks of the file that changed.

Comment by xorcist 1 day ago

"Tiny" speed improvement? According to the benchmarks, it's almost 10x the speed of rsync!

Really serious benchmarks never contain any information about network characteristics, options used or buffer sizes, and these are indeed very serious. The "details" button produces huge bar graphs. No disappointment there.

Absolutely looking forward to 10x my speed, perhaps even transferring gigabytes per second on my gigabit link.

Comment by xz18r 2 days ago

You live up to your username!

Comment by Zenul_Abidin 1 day ago

Wow, interesting! How is the performance compared to rclone though?

Comment by greaber 17 hours ago

I did some benchmarks against rclone today. https://greaber.github.io/syq-bench/rclone.html Syq was sometimes much faster and never more than a few percent slower. I didn't do any tuning of syq since it is meant to perform well without any special options, but I did tune a couple of rclone options. It's possible that more extensive rclone tuning would yield better results though. Please let me know if there is anything I need to try!

Comment by mika6996 2 days ago

Why is this better than rsync?

Comment by greaber 2 days ago

There are a bunch of reasons, some of which are explained in the description and in the docs. For me, the most important benefit is speed of copying. Note that syq doesn't currently implement rsync's delta merge algorithm, which allows it to avoid copying data that is already in the destination file but at a shifted offset. I might implement this (or an enhanced version of it) in the future. If your workloads have a lot of cases like this then syq might not be better than rsync for you, but I found that for me this rarely came up.

Comment by craftkiller 2 days ago

One instance where rsync's delta merge comes up for me: For all of my virtual machines, I build a bootable live ISO so that way I can cleanly swap out the full ISO with a new one when there are software updates. This means that every time I want to change a config file inside the virtual machine, I end up building a new ISO and rsyncing it to my server. rsync speeds this up significantly since most (but not all) of the ISO is unchanged.

Comment by greaber 1 day ago

Nice example. Rsync might be better here. Although syq will still avoid copying most of the image if changing the config file just modifies a small number of locations and doesn't globally move data around (like prepending a single byte to the ISO)

Comment by dwedge 2 days ago

How do you handle updates this way? Do you mount the iso, update and keep it as an iso? Are all of the virtual machines mostly read only?

Comment by craftkiller 1 day ago

The iso itself is read-only. When I need to update software I build a whole new iso and replace the existing one. For any folders/files that need persistence across reboots, I store that data separately from the iso, either in a virtual nvme drive or a mounted 9pfs folder from the host. The iso mounts the persistence drive/9pfs via /etc/fstab. So in the end, the software and config files are read-only (until I replace them with a new iso) but my databases and data files are read-write.

Getting into the unnecessary details: in classic live ISO fashion, the root (/) filesystem is tmpfs. Then I have at most 1 persistence drive and/or at most 1 9pfs mount. From there, I use bind mounts to only persist the few folders/files that I want preserved. For example, I'll have my persistence drive mounted at /persist and I'll bind mount "/var/lib/etcd" to "/persist/var/lib/etcd". Then when I reboot, everything outside of "/persist" is wiped out and I'm back to the initial iso + anything in /persist. So it is very similar to running a docker image with a volume mount, or running tails with persistence.

I build it all via NixOS and I use Impermanence to manage the bind mounts to the persist drive: https://wiki.nixos.org/wiki/Impermanence

Comment by amelius 2 days ago

There's something vastly superior over rsync still: "btrfs send" and "btrfs receive" (but require you to have btrfs on source and destination).

Comment by greaber 2 days ago

Thanks for the suggestion! I'm not very familiar with btrfs send and btrfs receive. Browsing the documentation, it looks cool, but maybe you could say more about the specific things that make it so much better for your use case? Could any of them be incorporated into a more general tool like syq? Also, it doesn't look like btrfs send helps transfers go fast by parallelizing them or has anything like syq's remote-remote transfer feature, which made me wonder if it would actually make sense to make syq able to work as an authenticated fast transport for a stream generated by btrfs send/receive.

Comment by gcr 1 day ago

'btrfs send' is more like diff/patch than it is like rsync. It requires both source and destination to agree about the source snapshot.

Comment by amelius 1 day ago

btrfs send is useful for sending incremental updates; this is used a LOT for backups; and it saves a lot of time because it does not need to traverse files/directories that were not changed

You cannot do this without help from the filesystem. That's why rsync itself does not support this, and it has to be done by a specialized btrfs tool in this case.

Comment by Someone 2 days ago

> Note that syq doesn't currently implement rsync's delta merge algorithm

That, and still claiming it’s “better than rsync”? Seems like you shouldn’t compare it with rsync (yet, if doing delta merges is planned)

Comment by greaber 1 day ago

Yeah, I kind of regret writing "better than rsync" in the tagline. Maybe I should have written "better than rsync in some respects".

Comment by nimih 2 days ago

Browsing the repo, I can see a couple of advantages:

- Single contributor. Software bugs are generally caused by developers writing code. By reducing the number of contributors, syq has cleverly reduced the surface area for defects to sneak in.

- Distribution via shell script. rsync is bundled in most linux distributions, which means you have to deal with annoying software updates from time to time. Syq, OTOH, tells you to pipe curl into bash to execute a shell script, which means one-and-done installation and maintenance.

- UI clarity. If you search StackOverflow for rsync, you'll see thousands of questions asking how to accomplish various tasks with rsync, both straightforward and arcane. By contrast, `syq` isn't even a tag on SO. The obvious conclusion here is that rsync's interface is so byzantine, and its documentation so poor, that users must resort to asking strangers for help, a problem obviously not shared by syq.

Comment by greaber 2 days ago

- No need to be snarky. It's a new project, while rsync is time tested, and that is certainly an advantage for rsync and all old software.

- You can install by piping curl to bash, using brew, or by compiling yourself. One thing about distribution I would argue syq gets right is that you are never relying on whatever version of syq happens to be installed on a server. The client syq always talks to a server syq tagged with the exact same version, and when your local copy of syq installs a remote copy, it verifies that the remote binary is signed by me.

- I never said or implied that rsync has a byzantine interface or poor documentation. I think rsync is well documented and the interface is overall fine. Syq even has an rsync compatibility mode, so you don't have to learn any new syntax if you don't want. Still, I tried to provide good docs for syq too, and honestly even if there were a gap in the docs, all you have to do these days is ask AI to look at the source code and tell you what to do. I think this also mitigates the trust issue with a new project from a single contributor.

Comment by nimih 1 day ago

FWIW, I'm only half-joking with my comment: dealing with software updates is an annoying part of life, and IMO rsync's UI is rather complicated (in that it's easy to use the wrong set of flags and do the wrong thing, sometimes in annoyingly subtle ways) and I basically always need to refer to the man pages (or, these days, an LLM, followed closely by the man pages to sanity-check what I'm doing) whenever I'm writing a new incantation of it rather than copying from an existing script or my shell history. I just chafe a little bit at declarations like "better than," since that's clearly not true (in the same way that rsync is not unequivocally "better than" syq), when more precise and less conceited descriptors like "faster than" are sitting right there on your benchmarks page.

Comment by greaber 1 day ago

Yeah, I get that. I really didn't mean to sound conceited, and I don't think that syq is better than rsync for every use case. I was just trying to convey in a very short space that it is something that rsync users might be interested in because it solves some frustrations with rsync.

Comment by ProphetOfParado 2 days ago

How do you hope to convince sys-admins to install this? I work with clusters where I do not have superuser access, so this is a no go for me.

Comment by greaber 2 days ago

It just installs in $HOME/.local/bin by default (and some other parts go in other locations inside your home directory). You don't need superuser access for anything. The first time you connect to a server, it installs its matching counterpart in your home directory on the server.

Comment by ProphetOfParado 1 day ago

Ah, that's nice!

Comment by unsnap_biceps 2 days ago

Honestly, that's even scarier for an untrusted/low trusted tool...

Comment by cstrahan 1 day ago

What is it about a tool, running as your own non-root user, installed in your home directory, that makes it scary?

Comment by unsnap_biceps 1 day ago

Let's say this is widely used and let's presume that a specific version has a security issue. With everyone having their own copy of the binary on every host versioned at the time they first ran a command on that host, or any decently sized fleet, it'll take a big effort to track down and ensure all versions have been updated or removed (to be re-generated later).

Frankly, we would re-kick our fleet ahead of schedule rather then try to remediate it more manually. But we re-kick our fleet on a rotating yearly schedule, so it's not too much an effort to re-kick it earlier. It's a completely automated system in place now.

Comment by throw1234567891 1 day ago

If I wanted to know what chat gpt "thinks", I'd ask chat gpt.

Comment by hikarudo 2 days ago

Very cool!

How does the direct TCP mode work?

Comment by greaber 2 days ago

Currently, there is just a fixed range of ports that it will use if they are open and you don't pass `--no-tcp`. It does its own encryption over the TCP connections. Also worth knowing is that for some connections (e.g. if there are any dropped packets), you may get much better performance if you can enable BBR congestion control. https://greaber.github.io/syq/server-tuning.html#test-conges...

Comment by QuiCasseRien 1 day ago

rclone has been the real fully featured ultra power alternative to rsync over the past 4-5 years.

How do you compare Syq to rclone ?

Comment by greaber 17 hours ago

As for speed, see https://greaber.github.io/syq-bench/rclone.html. As for other stuff, it depends what you use rclone for.