Development log: Minimizing S3 transfer during sync

Apr 8, 2020

Over the years my AWS bill has been slowly creeping up. Nothing serious – now in the $10s/month where it started in the $1s – but I’ve had a sneaking suspicion for a while that it’s not because my monthly visitor numbers have improved that much.

Popping open AWS’ Cost Explorer, I wasn’t surprised to find that S3 was the main culprit, but was surprised to find that the bulk of my costs weren’t monthly retention or data transfer out, but data transferred in.

This seemed somewhat incredible, so I started examining the details of the builds of this website as it runs its CI from GitHub Actions and auto-deploys to S3. The basic process involves (1) compiling the Go program from source, (2) having it build the static site, then (3) using the AWS CLI, specifically aws s3 sync, to send those to S3:

aws s3 sync ... s3://...

It seemed to be working correctly, except that on closer inspection of the last step, for a command called sync, there sure wasn’t a lot of syncing going on. Every file was being copied to S3 every time.

In some places that made sense. aws s3 sync figures out what needs to be resent based on the modification times (mtimes) of a local and remote target sharing the same name. HTML sources are built from scratch on every CI run, so it’s not surprising that they’re sent every time. However, a large bulk of the site’s total size are static assets and those are not rebuilt, and rarely even modified. Once pushed to S3, they should be fresh for a long time to come.

So what was going on? Well as it turns out, Git. Git preserves a large amount of data on every file that comes into its graph – contents, author, commit date and context, but quite notably, doesn’t preserve any file’s modification time. Each of my CI runs was doing a fresh git clone from source and every static asset was getting the current time as its last modification date, thereby making every file in my S3 bucket stale. I was pushing 100+ MB of static assets to S3 on every build, and a cron job was triggering builds hourly, adding up to on the order of gigabytes per day. Again, it’s S3, so luckily that won’t break the bank, but it was still a lot of needless waste.

The most obvious mitigation would be to stop relying on modification times and use an alternative comparison to function figure out what to sync. If using rsync for example, the --checksum flag would tell it to use checksums for the comparison instead of mtime and size. Unfortunately though, aws s3 sync doesn’t support that or any other alternatives …

… vvvvvvvvvvvwwppppp ztűrÅ‘ tükörfúróg *RECORD SCRATCH*

Okay, full disclosure time. I wrote a custom Go program to solve the problem by checking the commit history of every image asset via git log, then using the timestamps of those commits to set stable mtimes for sync. It was a little slow, and a lot of a hammer, but it worked perfectly, saving me 100+ MB of S3 ingress on every build.

But as I was writing this article about it, I decided to go back and read the manual one more time. I’d done it before to verify that nothing like --checksum existed in the AWS CLI, but not carefully enough. On that second pass I discovered the --size-only option. aws s3 sync works by comparing mtimes and sizes, and although --size-only doesn’t fundamentally change that strategy, it does disable the mtime half of the equation – files are considered equivalent as long as their name and size are equal.

--size-only has some risk for text-based content where minor changes (e.g. one character changed) might lead to a file of the same size. But for media assets where even a minor tweak tends to rewrite many bytes, it’s fine, and very fast.

I threw out my custom script solution and replaced it with --size-only, which also worked perfectly, and without all the custom code. I decided to publish this article anyway as a reminder to not be too quick to jump into writing new software – sometimes the right answer is to read more carefully.

Did I make a mistake? Please consider sending a pull request.