write hedging in Amazon DynamoDB
“Write hedging” is a nice technique to address p99 tail latencies, by increasing the volume of writes (or in the case of read hedging, reads):
Imagine you want a very low p99 read latency. One way to lower tail latencies is to hedge requests. You make a read request and then, if the response doesn’t come back quickly enough, make a second equivalent hedging request and let the two race. First response wins. If the first request suffered a dropped network packet, the second request will probably win. If things are just temporarily slow somewhere, the first request will probably win. Either way, hedging helps improve the p99 metrics, at the cost of some extra read requests.
Write hedging has a little more complexity involved, since you want to avoid accidental overwrites during races; this blog post goes into some detail on a technique to do this in DynamoDB, using timestamps. Good stuff.
(via Last Week In AWS)
Tags: via:lwia aws dynamodb write-hedging read-hedging p99 latencies tail-latencies optimization performance algorithms
Archives
-
I’m pretty happy with my current setup for the home network, but this is one I’ll keep in the back pocket for future possible use:
[Tailscale Docker Proxy] simplifies the process of securely exposing services and Docker containers to your Tailscale network by automatically creating Tailscale machines for each tagged container. This allows services to be accessible via unique, secure URLs without the need for complex configurations or additional Tailscale containers.
Tags: docker tailscale containers home networking