Google publishes details of internal database Spanner
New paper reveals details of next-generation database system, which is underpinned by excellent timekeeping.
company dedicated to making world’s information available, Google
are surprisingly secretive about their own backend. However, this
week they shed a small amount of light on some of the technology
that powers their infrastructure, a database system called
Detailed in a technical paper, which is to be officially presented at OSDI (Operating System Design and Implementation) conference in Hollywood next month, Spanner is described as “a scalable, globally-distributed database [...] designed to scale up to millions of machines across hundreds of datacenters and trillions of database rows”.
Managing huge amounts of data is obviously crucial to Google’s business plan; after all, with 425 million Gmail users – not to mention YouTube, Google+, and their other myriad services – existing options generally aren’t good enough. In the paper, the authors report issues that have arisen with use of BigTable and Megastore in existing projects, and how they have spent five years building their own database system to power F1, their advertising backend.
Spanner is able to provide high availability “even in the face of wide-area natural disasters” by replicating data “within or even across continents.” Its TrueTime API, dependent on GPS and atomic clocks to keep perfect time, is described as the “linchpin” for two unique features of Spanner: replication configurations can be controlled dynamically by applications, to control read and write latency; and providing externally consistent reads and writes.
The paper’s 26 authors conclude:
We have shown that reifying clock uncertainty in the time API makes it possible to build distributed systems with much stronger time semantics. In addition, as the underlying system enforces tighter bounds on clock uncertainty, the overhead of the stronger semantics decreases. As a community, we should no longer depend on loosely synchronized clocks and weak time APIs in designing distributed algorithms.
There’s some interesting lessons to learn here, particularly in
regards to data replication, which could easily be applied to other
systems. While Spanner’s source is still very much closed, the
paper still provides some interesting solutions to problems faced
by engineers everywhere.
For more info, ZDNet have a fairly good breakdown, or take a gander at the paper itself.
Photo by Les Chatfield.