I’ve spent some time over the last week writing amazonka-s3-streaming () which provides a conduit based streaming interface for uploading files/data to Amazon’a S3 service. There’s also a way to upload files or strict ByteStrings to S3 concurrently using the async package. The package is based on Brendan Hay’s amazing Amazonka () suite of libraries.
I’ve tried to make things as efficient as possible, and to avoid excess memory usage if I can. A few of the ways I’ve attempted to do this are:
mmaping file segments
when concurrently uploading a file, each file part (currently 6MiB) is read in using
mmap. The ensures that the mappings can be garbage collected as soon as they’re sent, usually resulting in much lower memory usage that mapping the whole file might.
The one big caveat here, which is unlikely to affect many people, but worth keeping in mind is that mmapping ByteStrings can break referential transparancy - if another process modifies the file, the contents of these parks might change.
Initially I was accumulating data in the conduit streaming interface using a ByteString
Builder, and then converting this to a Lazy ByteString for upload. It occured to me that
- the bytestrings are already in memory
- by converting them to Builders, buffers would be allocated to produce the Lazy ByteString
- these buffers would be immediately written to the socket
- so why not just use the already allocated ByteStrings and send them to the socket.
so I changed the implementation to store a (D)List of ByteStrings until more than 6MiB have been accumulated, and this list is then turned into a conduit
Source to be sent by the
UploadPart request to S3. I also keep track of the length and the SHA256 hash and total length, which are needed by
The Builder based implementation was my initial experiment, and didn’t survive long enough compare the performance with the list based implementation, so I don’t have benchmarks to prove this was a sensible move or not.
My experiments so far have shown that the concurrent upload is capable of saturating a 1Gbps connection quite easily, while using ~70MB RAM uploading a multi-hundred-megabyte file. The streaming interface has managed up to 40MB/s in my testing in GHCi. I’ll update this post once I’m back at work in the new year with some numbers.
I was inspired by a question by Alex Babkin about how to stream data into S3. His usecase is pretty cool, wanting to use his Qmuli package to stream data through AWS Lambdas written in Haskell without needing to download the entire file. Hopefully this package will allow this, while using no more than a few tens of MB more RAM than you would expect if you were able to stream the data directly over the network.
I’d love any feedback you’ve got, either here or on Reddit .