Finding a needle in Haystack Facebook’s Photo Storage

Shakthi Bachala

Outline

• Scenario• Goal• Problem• Previous Approach• Current Approach• Evaluation• Advantages• Critic• Conclusion

Scenario :April 2009 October 2011

Total 15 billion Photos 4*15 billion images= 60 billion images1.5 petabytes of data

65 billion Photos4*65 billion images =260 billion images20 petabytes of data

Upload Rate 220 million photos / week25 terabytes of data

1 billion photos / week60 terabytes of data

Serving Rate 550,000 images / sec 1 million images / sec

Goal

• High throughput and low latency• Fault-tolerant• Cost-effective• Simple

Previous Approach : Typical design for Photo Sharing

Previous Approach : NFS based design for Photo Sharing at facebook

Previous Approach – NFS based design

• Traditional file system architecture performs poorly under Facebook's kind of workload

• NFS - based Design: CDN effectively serves the hottest photos (profile pictures and recently updated photos), but facebook also generates a lot of requests for less popular images (long tail images). These are not handled by CDN

• Normal website had 99% CDN hit rate but facebook had around 80% CDN hit rate

8

Long Tail Issue

Previous Approach cont..

Problems with that approach were:• Wastage of storage capacity due to metadata

– Large metadata per file– Each image stored as a file

• Large number of disk operations for reads– Because of large directories (large directories containing thousands of

files)– Change of the directory structures and changing from large directories

to small directories has brought down the iops approximately from 10 to 2.5-3.0

Current Approach – Haystack Architecture

Current Approach- Haystack Components

The main components of Haystack architecture are:1. Haystack Directory2. Haystack Cache3. Haystack Store

Current Approach- Haystack Directory

The main goals of directory are:• Map logical volumes to physical volumes

– 3 Physical volumes( on 3 nodes) per one logical volume• Load balance

– Writes across logical volumes – Reads across physical volumes (any of the 3 stores)

• Caching strategy: Whether the photo request should be handled by the CDN or by the cache– URL generation

http://<CDN>/<Cache>/<Node>/<Logical volume id, Image id>

• The directory would Identify the logical volumes that are read only either because of operational reason or because those volumes have reached their storage capacity

Current Approach- Haystack Cache

• Approach:– The Cache receives HTTP requests for photos from

browser or CDNs– It is a distributed hash table with photo id as the key to

locate the cached data– If the photo id is missing in cache , the cache fetches the

data from photo server and replies it to the browser or CDN depending on the request

Current Approach- Haystack CacheCaches a photo if it satisfies the following two conditions:

• The request directly come from a user and instead of CDN– Facebook’s experience with the NFS-based design showed post-CDN caching

is ineffective as it is unlikely that a request misses in the CDN would hit in our internal cache

• The photos is fetched by the write enabled store– Photos are most heavily accessed soon after they are uploaded – File systems generally work better when doing either writes or reads but not

both

Current Approach- Haystack Cache Hit Rate

Current Approach : Haystack Store

• Replaces the storage and photo server layer in NFS based Design with this structure:

Current Approach : Haystack Store

• Storage : – 12x 1TB SATA, RAID6

• Filesytem:– Single approx. 10 TB xfs filesystem.

• Haystack:– Log structured , append only object store containing

needles as object abstractions– 100 haystacks per node each 100GB in size

Current Approach: Haystack Store File

Current Approach: Operations in Haystack

• Photo Read– Look up offset /size of the image in the incore index– Read Data (approx. 1 iop)

• Photo Write– Asynchronously append images one by one to the haystack

file– Next haystack file when becomes full– Asynchronously append index records to the index file– Flush index file if too many dirty index records– Update incore index

Current Approach: Operations in Haystack

• Photo Delete– Lookup offset of the image in the incore index– Mark the image needle flag as “DELETED”– Update incore index

• Index File:– Provides minimum metadata to locate the needle in the

Haystack store– Subset of Header metadata

Current Approach: Haystack Index File

Haystack Based Design - Photo Upload

Haystack Based Design - Photo Download

Current Approach: Operations in Haystack

• Filesystem:– Haystack uses XFS, an extent based file system

• It has two main advantages:– The block maps for several contiguous large files can be

small enough to be stored in the main memory– XFS provides efficient file pre allocation, mitigating

fragmentation and reigning in how large block maps can grow

Current Approach: Haystack Optimization

• Compaction:– Infrequent online operation– Create a copy of haystack skipping duplicates and deleted

photos– The patterns of deletes to photo views, young photos are a

lot more likely to be deleted– Last year about 25% of the photos got deleted

Current Approach: Haystack Optimization

• Saving More Memory:– With the following two techniques store machines reduced

their main memory footprints by 20%– Eliminate the need for an in-memory representation of

flags by setting the offset to be 0 for deleted photos.– Store machine do not keep track of cookie values in main

memory and instead check the supplied cookie after reading from the disk

Current Approach: Haystack Optimization

• Batch Uploads:– Disks perform better with large sequential writes instead

of small random writes, so facebook uses batch uploads whenever possible

– Many users upload entire albums to facebook instead of each picture which gives an opportunity to batch the uploads

Evaluation -Data

Evaluation – Production Workload

Advantages

• Simple design• Decrease number of disk operations by

reducing the average metadata per photo• This system is robust enough to handle a very

large amount of data• Fault Tolerant

Critic

• I thought this approach is very facebook specific .

• Any other?

Conclusion

• Built a simple but robust data storage mechanism for facebook photo storage to accommodate long tail of photo requests which was not possible by previous approaches

References

1. http://www.usenix.org/event/osdi10/tech/2. http://

www.cs.uiuc.edu/class/sp11/cs525/sched.htm