ORLANDO, Fla. -- Storage administrators spent time this week at Storage Networking World investigating object storage issues, such as how to handle migration of metadata from legacy storage systems, whether erasure coding
One object-based storage customer demonstrated how he overcame the hurdles of migrating from legacy network attached storage (NAS) to object storage.
Justin Stottlemyer, fellow and storage architect for the cloud photo-sharing website Shutterfly, built a primary archive that consists of 20 billion photo images and 80 petabytes (PB) of capacity on 400 Cleversafe object storage-based Slicestor 1420 storage nodes fronted with 60 Cleversafe Accesser 2100 appliances that communicate via HTTP or REST to Web servers.
"Migrating [from NAS] is simpler than you think. For us, it was as simple as plug it in, make a few application tweaks and let it run," said Stottlemyer, who designed Cleversafe's 1420 hardware device and now is a technical advisor for the object storage company. "The goal was to take all the new ingested data and enable it to write to object stores. We also moved data from legacy file systems to Cleversafe object storage."
When Stottlemyer arrived at Shutterfly, it had 20 PB of capacity on hundreds of PolyServe file systems and 20 storage arrays from DataDirect Networks and Hitachi Data Systems. The new object architecture was rolled out within a few months after a six-week, two-site production test. Storage costs have been cut by 50%, he said.
The main adjustments included recoding applications to use PUT and GET commands instead of reads and writes of data. A second database was connected to the upload application server that was designed with an extra column for global unique object IDs, a change that came in handy when Shutterfly had to migrate 8 PB of storage into the primary storage archive after acquiring Kodak Gallery in May 2012 for $23.8 million.
"We did it in 82 days," said Stottlemyer, who worked at Facebook, eBay and PayPal before moving to Shutterfly. "We created a small database for this. We pulled in all the file system metadata and added an extra [database] column for object IDs, then split the workload into small chunks and copied them into object stores."
RAID 5 data protection was replaced by Cleversafe's erasure coding, a move that Stottlemyer considered key after the company experienced a simultaneous failure of 178 hard drives that handled about 2.2 PB of data within months of his taking the job. The drive failure left LUNs without parity protection during the five days it took to do a recovery.
"[RAID] stops scaling at the multiple petabyte level," Stottlemyer said. "With 2 [TB], 3 TB and 4 TB drives, time to generate parity grew from days to weeks."
Not everyone is sold on object storage yet, however. Dennis Ballagh, manager of solution architecture at The Walt Disney Company, said it's unlikely he would use object storage without a traditional file system layer.
"The only way I would consider using object storage is with a file system in front of it," Ballagh said. "I don't have applications that can take advantage of the API [application programming interface] calls. There are benefits of getting the file system out of the way so applications can put and get data directly from object storage because applications don't need to be location aware. But we are not there yet."
Ballagh's not ready to replace RAID with erasure coding yet, either. "What's the business need?" he said.
Shutterfly's Stottlemyer said companies should consider how users and the application use storage when they design an object storage-based architecture. "Are the blocks large or small? Is the data traffic random or sequential? Is your company a B2B or B2C? Does it experience weekly, monthly or seasonal traffic patterns?
"The idea is to take big risks intelligently," Stottlemyer said.