F1: A Distributed SQL Database That Scales

http://research.google.com/pubs/pub41344.html

by alec • 13 years ago

131 23 13 years ago

research.google.com

dvliman• 13 years ago

This is not new by any mean...

jahewson• 13 years ago

The paper was only published in conference proceedings yesterday.

paulsamways• 13 years ago

See http://research.google.com/pubs/pub38125.html

packetslave• 13 years ago

"not new" != "not interesting"

jacques_chester6• 13 years ago

Impressive, as one might expect. Some reactions:

-- Secret sauce

A lot of the magic of F1 comes from Spanner, the distributed storage system. The name "F1" itself is an allusion to "inheriting" some of the properties of Spanner.

-- Hierarchical tables

What they call hierarchical tables, I would think best be viewed as one-to-many relationships. In guess they've privileged this model in their storage because that's what a lot of their AdWords schema looks like.

-- Change History

I like the observation that keeping full histories is relatively straightforward with atomic, granular timestamping and indeed that it should be baked in. Every database schema I've ever worked with always goes through a similar evolutionary cycle:

1. We only need to capture the current state of the model.

2. Wait, we do need to capture historical states of the model.

3. Wait, the model is changed, we need to capture historical states and the models that were current.

(You can think of this as taking progressive differentials of incoming transactions).

The F1 designers have baked that right into the database, where it belongs. Weak temporal support has long been the sore point in SQL.

-- Remote data

I was struck by their observation that most database storage engines are built around the concept of seeks and reads, whereas theirs is necessarily built around batching and pipelining over a network. If I am reading them correctly, their engine takes advantage of having multiple copies of data by sending reads to multiple disk nodes and then working from the first copy that is returned.

vosper• 13 years ago

Your point about change history and its poor support in SQL is spot-on; at my employer we've gone through (1) and (2) and are trying to figure out (3) at the moment.

From what I understand of Datomic temporal support is baked in, but it's about the only database I can think of that does that, and I haven't heard stories about it being used in production for large volumes of data.

jacques_chester6• 13 years ago

Snodgrass wrote an entire book about dealing with the temporal blindness of SQL which might help you:

http://www.cs.arizona.edu/people/rts/tdbbook.pdf

More generally, you can usually build a model that handles (2), the changing state, by generalising your original model. I've done in both of classical ways, having either validity period fields (thus pushing complexity into every query), or by having audit tables (thus relying on triggers).

But building a changing model is hard, because at the schema level, SQL only supports (1). That is, it provides primitives to change the current state of the model.

So you wind up having to build a meta-model. The deficiencies of SQL (and I'm an RDBMS bigot) mean that we wind up building inner platforms. Or outer platforms -- witness migration toolkits.

Yet support for changing models is essential. No model is ever constant. Most of my work is for a public-sector client whose legal reporting requirements are constantly changing. You cannot throw away old reports which were made under old regulations; you must be able to recreate the world at any point in time. Which requires either a lot of bookkeeping code or taking periodic snapshots.

Datomic looks neat. I see that I'm not the only one who made the leap from transactional memory to transactional models.

Edit: I just found notes I wrote in 2009 on database languages while looking for something else. Spooky: http://chester.id.au/2013/08/28/notes-towards-a-set-objectiv...

lucian1900• 13 years ago

Also, Datomic is not open source :(

mmastrac• 13 years ago

[I] guess they've privileged this model in their storage because that's what a lot of their AdWords schema looks like.

This is actually the model they've used in things like AppEngine as well - nested entities allow them to take advantage of locality in transactions.

drunkpotato• 13 years ago

How do hierarchical tables compare to Postgres' hstore?

I think recursive structures like lists and trees are one weakness of the relational model, and haven't found a fully satisfying relational answer to the limitations. It seems to me like the assumption of atomic column types is a major weakness inherent in the relational model itself when it comes to recursion. Any thoughts/comments?

jacques_chester6• 13 years ago

How do hierarchical tables compare to Postgres' hstore?

If I am reading them correctly, it's a storage strategy, not a specific "feature" per se. The closest analogy to hstore is that they provide native support for storing and querying protobuf blobs.

I think recursive structures like lists and trees are one weakness of the relational model, and haven't found a fully satisfying relational answer to the limitations.

It depends on why you're using trees or graphs.

If it's inherent in the data, then modern SQL has recursive queries that make it much easier than the old methods.

If it's inherent in the model, you will find it harder. You might need to pick a non-standard approach, such as PostgreSQL's inherited tables. I'd think long and hard before saying it's inherent in the model, by the way. Strictly speaking you can represent the same thing as sets of relations or as a graph; it's better to utilise the strengths of the tool in front of you.

zapov2• 13 years ago

Both Oracle and Postgres have objects and collections. This allows you to build hierarchical structures. Postgres unfortunately doesn't allow for recursive structures ;(

If you are interested in how to do it take a look here: https://blog.dsl-platform.com/postgres-bridge-between-worlds...

fintler2• 13 years ago

I posted this question on StackOverflow regarding TrueTime (used by Spanner) a few days ago and haven't received any responses:

<http://stackoverflow.com/questions/18384883/why-is-googles-t...

However, I think this HN thread seems like it might be a good place to get comments on why my line of reasoning may be incorrect. Does anyone have any thoughts on why building something like Spanner on top of basic Paxos quorums and NTP would be a bad idea?

mad44• 13 years ago

Check this out. http://muratbuffalo.blogspot.com/2013/08/beyond-truetime-usi...

fintler2• 13 years ago

Thanks for the link -- I'm working through it now.

capkutay• 13 years ago

Can someone enlighten me as to why google doesn't want to make an enterprise play? Many companies tout that they are offering data infrastructures similar to those used at google..why wouldn't google commercialize that technology themselves? It should be pretty obvious as to how that could be of huge financial benefit to the company.

dragonwriter2• 13 years ago

Can someone enlighten me as to why google doesn't want to make an enterprise play?

They do. They have a range of commercial, including enterprise, offerings (many of which are based on originally-internal technologies, including their MySQL-based distributed database that preceded F1.)

why wouldn't google commercialize that technology themselves?

They probably will, just as they have commercialized their previous internal storage technologies.

F1: A Distributed SQL Database That Scales

Related Stories