AWS at
Common Crawl

Stephen Merity

We're a non profit that makes

web data

accessible to anyone

Each crawl is billions of pages...


August crawl was

2.8 billion web pages

~200 terabytes uncompressed

Released

totally free


Lives on S3 (Public Data Sets)

Common Crawl Use Cases


Large scale web analysis:

Natural language processing:




We're the largest public user of

Trivia Time

What major OSS project grew out of Nutch?

Trivia Time

What major OSS project grew out of Nutch?


Crawling Architecture Overview


Standard Hadoop Cluster:
 On-demand: One master
 Spot Instances: ~60 m2.2xlarge or ~110 m1.xlarge

Generate, Crawl, Export (raw + processed)
 2-3 billion pages in 15-20 days


Hadoop is our hammer, handles all the nails
 (even if that nail is more like a screw)

Crawling Architecture Overview


Why those instances?

m1.xlarge
  • Best all round deal + usually cheap
    (1Gbps network, 1.6TB disk, 15GB RAM, ...)


m2.2xlarge
  • Similar specs to m1.xlarge, less bang for the buck


(Note: ec2instances.info is insanely useful)

Topics for Advanced AWS




  • Random access archives using S3

  • Surviving & thriving w/ spot instances

  • EMR - the good, the bad, the ugly

Random Access Archives

In an optimal world,
everything would be in RAM


For one dollar an hour...
  • EC2 RAM      = 87 GB
  • EC2 Disk        = 10,685 GB
  • S3 (standard) = 24,000 GB
  • S3 (reduced)  = 31,000 GB

How do we get advantages of RAM w/ S3?

Random Access Archives


gzip spec allows for gzip files to be stuck together

gzip + gzip + ... + gzip + gzip

Why are we interested?

☑ Re-uses existing data format + tools

☑ Advantages of per object compression (10% larger than full gzip)

☑ Partition into optimal sized collections (Hadoop / S3)

☑ Allow random access to individual objects

URL index


Maps a URL to its location in the archive using
Prefix B-Tree + S3 byte range


5 billion unique URLs x [ len(URL) + len(pointer) ]
= 437 GB
= $13 per month


Topics for Advanced AWS




  • Random access archives using S3

  • Surviving & thriving w/ spot instances

  • EMR - the good, the bad, the ugly

Spot Instances

Spot instances are vital but are problematic...
  • Availability issues
  • Cost fluctuations
  • Transient failure

Spot Instances


To survive transient failures trivially:
rely on S3 and not HDFS


HDFS is brilliant if disks fail rarely + independently


Solution:
Small jobs + push data to S3 on completion

Spot Instances


If you do need to keep state on HDFS,
rack awareness is vital

Segment your cluster's machines according to:
  • Region (minimize transfer costs)
  • On-demand versus spot instance
  • Machine type (prices mostly independent)
  • Spot prices

Topics for Advanced AWS




  • Random access archives using S3

  • Surviving & thriving w/ spot instances

  • EMR - the good, the bad, the ugly

What good is data if it's not used?

 

Accessing & using it needs to be easy

Elastic MapReduce

= Trading money for simplicity


It's far cheaper to spin up your own cluster... 

(✖ 100-200% overhead on spot instances)


m1.xlarge = $0.088 for EMR + $0.040 for spot


For contributors, we're happy to pay however:
time & energy on task and not cluster

Ganglia


EMR in Education


Universities should teach big data courses with ...

big data

English Wikipedia

= 9.85 GB compressed
= 44 GB uncompressed
If "big data" isn't bigger than your phone's storage ...

Python + mrjob

https://github.com/commoncrawl/cc-mrjob


☑ Test and develop locally with minimal setup 
(no Hadoop installation required)

☑  Only uses Python, not mixing other languages
(benefits of Hadoop ecosystem without touching it)

☑  Spin up a cluster in one click
(EMR integration = single click cluster)

Summary



  • Random access archives using S3
    Allows effective storage of enormous datasets

  • Surviving & thriving w/ spot instances
    S3 defends us against transient failures

  • EMR - the good, the bad, the ugly
    Simplifies and lowers barriers to entry

Interested in helping out on something amazing?
Send me a message and/or visit commoncrawl.org



Stephen Merity
smerity.com /  @smerity
@commoncrawl

Attributions


http://thenounproject.com/term/broken-computer/20293/

AWS at
Common Crawl

Stephen Merity

Made with Slides.com