Facebook dumps AWS, moves Instagram's images to own bit barns
Server count down by a third, says engineer, and Iowa data centre coming online now
Facebook has migrated Instagram's colossal collection of images out of Amazon Web Services and into its own bit barns.
News of the move comes in this interview with Facebook infrastructure engineer and Open Compute Foundation program development chap Charlie Manese conducted by Australian hyperscale compute appliance startup Infrx .
At about the 7:45 mark of the video, Manese says Instagram has moved onto Facebook infrastructure and now uses about a third fewer servers, “obviously reducing our costs from when we had it on the Amazon platform.”
Instagram accumulates 45 million new photos a day, he said, but that's dwarfed by the 300 million or so uploaded to Facebook each day. The Social NetworkTM also has handles six billion Likes and needs to make over 200 billion connections to ensure everything its users do show up on friends' timelines.
Menese also reveals that Facebook is “just starting to open” a new data centre in Iowa, adding to its Sweden facility and two others in the land of the free. The new Iowa facility is powered largely by wind, at a cost close to that of conventional energy sources.
Another nugget Menese reveals is that Facebook's Open Compute designs don't always deliver the same savings. Different workloads have different demands, he explains, so while power consumption is down everywhere, not everything Facebook does has generated the same 24 per cent server count reduction the company enjoys in its web server fleet.
Having said that, he also says “If we had to buy five or six times more servers I don't think Facebook would be where it is today”.
Facebook has doubtless built a rig tuned to Instagram's needs, so it is probably not reasonable to draw a conclusion that AWS can't keep up with the house of Zuckerberg's data centre skills. Having said that, AWS can't have enjoyed losing such a colossal client. That Facebook promotes Open Compute kit as a way to achieve hyperscale, on-premises or in the cloud, must also rankle as its open source designs create an ever-more compelling alternative. ®