Slicer: Auto-Sharding for Datacenter Applications

https://www.usenix.org/conference/osdi16/technical-sessions/presentation/adya

by mlerner • 5 years ago

58 12 5 years ago

gnabgib• 5 years ago

From 2016. It's nice to have the option between the slides[0], 22m of audio[1] and whitepaper[2]

[0] https://www.usenix.org/system/files/conference/osdi16/osdi16... [1]: https://2459d6dc103cb5933875-c0245c5c937c5dedcca3f1764ecc9b2... [2]: https://www.usenix.org/system/files/conference/osdi16/osdi16...

bognition• 5 years ago

Super interesting.

The TL;DR is that google built a system to shard client requests into different services/task by some largish shard key. The example they provide is to shard requests into ML model services by the underlying language (English vs Chinese) as different languages need different models.

Rather than requiring each product team to manage sharding, google built a custom layer to do it for any team to use. In practice it's not used by that many teams (~20 services at Google use it).

itcmcgrath• 5 years ago

* >20 when this was published in 2016.

jeffbee2• 5 years ago

Slicer is one of those things that enjoys a lot more currency outside Google than inside it, though.

new_realist• 5 years ago

Does Slicer work with range partitions?

jzelinskie2• 5 years ago

Slicer is mentioned briefly in the PWL talk on Zanzibar[0]. I've chatted with a few folks from Uber that worked on Ringpop[1], which was inspired by Slicer, but doesn't have any of the novel dynamic balancing properties. I still don't think there's anything anbalagous outside of Google.

[0] https://youtu.be/1nbSbe3kw2U?t=1612

[1] https://github.com/uber/ringpop-go

akshayshah• 5 years ago

I worked on ringpop and the associated tech at Uber, though I wasn’t part of the original design & implementation. This isn’t the origin story I remember. My recollection is that ringpop was inspired by the SWIM paper, Hashicorp’s Serf, and Microsoft Orleans, and was part of an overall desire to move away from Chubby-like CP systems to AP designs.

danpalmer• 5 years ago

Isn’t the sweet spot a combination of AP and CP for different parts of the same problem?

Arguably that’s what Slicer implements. It’s AP for high traffic routing requests that use it, and CP for slower decision making.

Same could be said for things like FoundationDB (and I’d guess Spanner) – topology in CP, “customer activity” in AP, resulting in a system that degrades very gracefully. A bit of a simplification, but I think it’s a good model for services like this.

jzelinskie2• 5 years ago

From the people that I've spoken to, you are correct that this is where the implementation eventually arrived, but I'm not sure that was the original intent. I know at one point they tried a Go rewrite, but I'm not sure that ever shipped or if any design changes were made.

Slicer: Auto-Sharding for Datacenter Applications

Related Stories