Benchmarking aiodynamo vs other libs

[homm]

Hi! This is friendly post with benchmark results which I have done for my library. I think it could be Interesting for you.

https://github.com/homm/aspen_dynamo#benchmark

Benchmark source code is available in the repo. All results for single get_item call. The main oddity is HTTPX results. Looks like it extremely slow compared to aiohttp.

[ojii]

Hi, thank you for your interest in this library and trying to benchmark it.

Some comments on aiodynamo performance goals (and to a lesser degree on your benchmark):

• aiodynamo is mostly optimized for throughput, not latency (get_load in your benchmark, much less get_latency)
• aiodynamo tries to provide good performance for mass reads (Query and Scan), as single-item operations GetItem, PutItem, UpdateItem, DeleteItem are generally fast enough and can usually be parallelised trivially)
• Last time I benchmarked this (which is many years ago at this point), aiodynamo had vastly superior item deserialization (DynamoDB JSON to Python, so {"S": "foo"} to "foo") performance than aiobotocore/boto3. This improvement was actually the largest performance gain for improving Query/Scan throughput of non-tiny items, which was the original reason to write this library in the first place.

The main oddity is HTTPX results. Looks like it extremely slow compared to aiohttp.

I thought I documented this (but looking at the docs I think I actually didn't), but the default aiohttp and httpx client configurations are not optimized for direct aiodynamo usage. For example, for aiohttp you really want a custom TCPConnector with a much higher limit. I don't have much experience in using httpx in projects that hit throughput/performance issues, but it may have similar client defaults that cause internal throttling.