Realtimeanalyticsattwitter strata2011-110204123031-phpapp02

Rainbird:
Real-time Analytics @Twitter
Kevin Weil -- @kevinweil
Product Lead for Revenue, Twitter

TM

Agenda
‣ Why Real-time Analytics?
‣ Rainbird and Cassandra
‣ Production Uses at Twitter
‣ Open Source

My Background
‣ Mathematics and Physics at Harvard, Physics at
Stanford
‣ Tropos Networks (city-wide wireless): mesh
routing algorithms, GBs of data
‣ Cooliris (web media): Hadoop and Pig for
analytics, TBs of data
‣ Twitter: Hadoop, Pig, HBase, Cassandra, data
viz, social graph analysis, soon to be PBs of data

My Background
‣ Mathematics and Physics at Harvard, Physics at
Stanford
‣ Tropos Networks (city-wide wireless): mesh
routing algorithms, GBs of data
‣ Cooliris (web media): Hadoop and Pig for
analytics, TBs of data
‣ Twitter: Hadoop, Pig, HBase, Cassandra, data
viz, social graph analysis, soon to be PBs of data
Now revenue products!

Why Real-time Analytics
‣ Twitter is real-time

Why Real-time Analytics
‣ Twitter is real-time
‣ ... even in space

Real-time Reporting
‣ Discussion around ad-based revenue model
‣ Help shape the conversation in real-time with
Promoted Tweets

Real-time Reporting
‣ Discussion around ad-based revenue model
‣ Help shape the conversation in real-time with
Promoted Tweets
‣ Realtime reporting
ties it all together

Requirements
‣ Extremely high write volume
‣ Needs to scale to 100,000s of WPS

Requirements

‣ High read volume
‣ Needs to scale to 10,000s of RPS

Requirements


‣ Horizontally scalable (reads, storage, etc)
‣ Needs to scale to 100+ TB

Requirements


‣ Horizontally scalable (reads, storage, etc)
‣ Needs to scale to 100+ TB

‣ Low latency
‣ Most reads <100 ms (esp. recent data)

Cassandra
‣ Pro: In-house expertise
‣ Pro: Open source Apache project
‣ Pro: Writes are extremely fast
‣ Pro: Horizontally scalable, low latency
‣ Pro: Other startup adoption (Digg, SimpleGeo)

Cassandra
‣ Pro: In-house expertise
‣ Pro: Open source Apache project
‣ Pro: Writes are extremely fast
‣ Pro: Horizontally scalable, low latency
‣ Pro: Other startup adoption (Digg, SimpleGeo)

‣ Con: It was really young (0.3a)

Cassandra
‣ Pro: Some dudes at Digg had already started
working on distributed atomic counters in
Cassandra

Cassandra
Cassandra
‣ Say hi to @kelvin

Cassandra
Cassandra
‣ And @lenn0x

Cassandra
Cassandra
‣ And @lenn0x
‣ A dude from
Sweden began helping: @skr

Cassandra
Cassandra
‣ And @lenn0x
‣ A dude from
Sweden began helping: @skr

‣ Now all at Twitter :)

Rainbird
‣ It counts things. Really quickly.
‣ Layers on top of the distributed
counters patch, CASSANDRA-1072

Rainbird
‣ It counts things. Really quickly.
‣ Layers on top of the distributed
counters patch, CASSANDRA-1072

‣ Relies on Zookeeper, Cassandra, Scribe, Thrift
‣ Written in Scala

Rainbird Design
‣ Aggregators
buffer for 1m
‣ Intelligent
flush to
Cassandra
‣ Query
servers read
once written
‣ 1m is
configurable

Rainbird Data Structures
struct Event
{
1: i32 timestamp,
2: string category,
3: list<string> key,
4: i64 value,
5: optional set<Property> properties,
6: optional map<Property, i64> propertiesWithCounts
}

struct Event
{ Unix timestamp of event
1: i32 timestamp,
2: string category,
4: i64 value,
}

struct Event
{ Stat category name
1: i32 timestamp,
2: string category,
4: i64 value,
}

struct Event
{ Stat keys (hierarchical)
1: i32 timestamp,
2: string category,
4: i64 value,
}

struct Event
{ Actual count (diff)
1: i32 timestamp,
2: string category,
4: i64 value,
}

struct Event
{ More later
1: i32 timestamp,
2: string category,
4: i64 value,
}

Hierarchical Aggregation
‣ Say we’re counting Promoted Tweet impressions
‣ category = pti
‣ keys = [advertiser_id, campaign_id, tweet_id]
‣ count = 1
‣ Rainbird automatically increments the count for
‣ [advertiser_id, campaign_id, tweet_id]
‣ [advertiser_id, campaign_id]
‣ [advertiser_id]
‣ Means fast queries over each level of hierarchy
‣ Configurable in rainbird.conf, or dynamically via ZK

‣ Another example: tracking URL shortener tweets/clicks
‣ full URL = http://music.amazon.com/some_really_long_path
‣ keys = [com, amazon, music, full URL]
‣ count = 1
‣ [com, amazon, music, full URL]
‣ [com, amazon, music]
‣ [com, amazon]
‣ [com]
‣ Means we can count clicks on full URLs
‣ And automatically aggregate over domains and subdomains!

‣ count = 1
‣ [com, amazon, music] How many people tweeted
‣ [com, amazon] full URL?
‣ [com]

‣ count = 1
‣ [com, amazon] any music.amazon.com URL?
‣ [com]

‣ count = 1
‣ [com, amazon] any amazon.com URL?
‣ [com]

‣ count = 1
‣ [com, amazon] any .com URL?
‣ [com]

Temporal Aggregation
‣ Rainbird also does (configurable) temporal
aggregation
‣ Each count is kept minutely, but also
denormalized hourly, daily, and all time
‣ Gives us quick counts at varying granularities
with no large scans at read time
‣ Trading storage for latency

Multiple Formulas
‣ So far we have talked about sums
‣ Could also store counts (1 for each event)
‣ ... which gives us a mean
‣ And sums of squares (count * count for each event)
‣ ... which gives us a standard deviation
‣ And min/max as well

‣ Configure this per-category in rainbird.conf

Rainbird
‣ Write 100,000s of events per second, each with
hierarchical structure
‣ Query with minutely granularity over any level of
the hierarchy, get back a time series
‣ Or query all time values
‣ Or query all time means, standard deviations
‣ Latency < 100ms

Production Uses
‣ It turns out we need to count things all the time
‣ As soon as we had this service, we started
finding all sorts of use cases for it
‣ Promoted Products
‣ Tweeted URLs, by domain/subdomain
‣ Per-user Tweet interactions (fav, RT, follow)
‣ Arbitrary terms in Tweets
‣ Clicks on t.co URLs

Use Cases
‣ Promoted Tweet Analytics

Each different metric is part
Production Uses of the key hierarchy


Uses the temporal
aggregation to quickly show
Production Uses different levels of granularity


Data can be historical, or
Production Uses from 60 seconds ago


Production Uses
‣ Internal Monitoring and Alerting

‣ We require operational reporting on all internal services
‣ Needs to be real-time, but also want longer-term
aggregates
‣ Hierarchical, too: [stat, datacenter, service, machine]

Production Uses
‣ Tweet Button Counts

‣ Tweet Button counts are requested many many
times each day from across the web
‣ Uses the all time field

Open Source?
‣ Yes! ... but not yet

Open Source?
‣ Relies on unreleased version of Cassandra

Open Source?
‣ ... but the counters patch is committed in trunk (0.8)

Open Source?
‣ ... also relies on some internal frameworks we need to
open source

Open Source?
open source
‣ It will happen

Open Source?
open source
‣ It will happen
‣ See http://github.com/twitter for proof of how much
Twitter open source

Team
‣ John Corwin (@johnxorz)
‣ Adam Samet (@damnitsamet)
‣ Johan Oskarsson (@skr)
‣ Kelvin Kakugawa (@kelvin)
‣ Chris Goffinet (@lenn0x)
‣ Steve Jiang (@sjiang)
‣ Kevin Weil (@kevinweil)

If You Only Remember One Slide...
‣ Rainbird is a distributed, high-volume counting
service built on top of Cassandra
‣ Write 100,000s events per second, query it with
hierarchy and multiple time granularities, returns
results in <100 ms
‣ Used by Twitter for multiple products internally,
including our Promoted Products, operational
monitoring and Tweet Button
‣ Will be open sourced so the community can use and
improve it!

Questions?
Follow me: @kevinweil

TM

Realtimeanalyticsattwitter strata2011-110204123031-phpapp02

More Related Content

What's hot (16)

Viewers also liked (7)

Similar to Realtimeanalyticsattwitter strata2011-110204123031-phpapp02 (20)

Recently uploaded (20)

Realtimeanalyticsattwitter strata2011-110204123031-phpapp02

Editor's Notes