Pachyderm (YC W15) Raises $2M to Power Big Data Processing for the Docker Era | Hacker News Reader

Pachyderm (YC W15) Raises $2M to Power Big Data Processing for the Docker Era

techcrunch.com

35 pointsjaz468 comments

Hi Guys, co-founder here. We'll be hanging out on this thread all day to answer any questions.

I'm curious about the decision to target Docker so specificially. Is that just how you plan to manage resources between different users / workloads? Or is there another benefit?

I work in the Hadoop space (although I like the stack as a whole, I do see the downside of having evolved the stack over time with a bunch of independent projects over time, so although I suspect Pachyderm will have a long way to go, I watch with interest). We mainly use Docker containers for testing / training, etc. A distributed system allows you to scale, but you still want each node to be as big as possible, so in production I'd never use Docker. YARN of course puts things in to "containers" to manage resources, but each container is still running in the same OS. Targeting a system that partitions the resources of a physical machine so completely seems counter-intuitive for Big Data, so what's the thinking there?

edit: or was the focus on Docker added by TC to attract more clicks? :)

Containers solve a few problems for us. They allow us to provide strong isolation guarantees for jobs and offer nice resource quota semantics (via cgroups.) They also serve as a good format for specifying jobs in because they can be easily passed around between nodes via docker pull and push.

Our use of Docker is very similar to how you describe YARN's use of containers. We deploy on a cluster of CoreOS machines and each machine is generally as big as possible. We then pack several containers on to that same machine so that we can get as much performance as possible out of it. This also gives us an opportunity to do smart things like sharing data volumes between docker containers that are running different workloads over the same data.

Could you please explain the difference between datomic (http://www.datomic.com/) and "The Pachyderm File System"?

Hi! Good question.

Datomic and pfs have very similar data models, we drew some inspiration for our architecture from them.

They're targeting different use cases though. Datomic is a database so it's meant as the source of truth that you run your application off of. It has features like optimized point queries, indexing, caching etc. Pachyderm on the other hand is meant for analytic workloads, so we focus on having very fast aggregations and tools for long running queries and workloads.

Hope this clears it up for you :)

Thanks for the answer, where can we submit our suggestions for examples/tutorials?

I mean having an example project that runs some Python/R/C# code against historical stock market data would be a good use case.

This example seems really interesting, we should do it.

Github issues would be a fine place I think.

Or emailing me at jd@pachyderm.io if you'd prefer :)

I'm sorry and don't want to be negative. Only helpful. Change your name asap.

You don't want "derm" on there. This is not a skin medicine.

Pachy is a little odd too; but less so.

Pachyd ?

Reply on news.ycombinator.com