Cassandra from the trenches: migrating Netflix (update)

Cassandra from the trenches:migrating Netflix

Jason BrownSenior Software Engineer

Netflix

@jasobrown jasedbrown@gmail.com

http://www.linkedin.com/in/jasedbrown

History, 2008

• In the beginning, there was the webapp– And a database– In one datacenter

• Then we grew, and grew, and grew– More databases, all conjoined– Database links, PL/SQL, Materialized views – Multi-Master replication (MMR)

• Then it melted down– Couldn’t ship DVDs for ~3 days

History, 2009

• Time to rethink everything– Abandon our datacenter– Ditch the monolithic webapp– Migrate single point of failure database to …

History, 2010

• SimpleDB/S3– Managed by Amazon, not us– Got us started with NoSQL in the cloud– Problems:• High latency, rate limiting (throttling)• (no) auto-sharding, no backups

Shiny new toy (2011)

• We switched to Cassandra– Similar to SimpleDB, with limits removed– Dynamo-model appealed to us– Column-based, key-value data model seemed

sufficient for most needs– Performance looked great (rudimentary tests)

Data Modeling - Where the rubber meets the road

About Netflix’s AB Testing

• Basic concepts– Test – An experiment where several competing

behaviors are implemented and compared– Cell – different experiences within a test that are

being compared against each other– Allocation – a customer-specific assignment to a

cell within a test

Data Modeling - background

• AB has two sets of data– metadata about tests – allocations

AB - allocations

• Single table to hold allocations– Currently at > 1 billion records– Plus indices!

• One record for every test that every customer is allocated into

• Unique constraint on customer/test

AB – relational model

• Typical parent-child table relationship• Not updated frequently, so service can cache

Data modeling in Cassandra

• Every where I looked, the Internet told me to understand my data use patterns

• Identify the questions that you need to answer from the data

• Know how to query your data set and make the persistence model match

Identifying the AB questions that need to be answered

• High traffic– get all allocations for a customer

• Low traffic – get count of customers in test/cell– find all customers in a test/cell– find all customers in a test who were added within

a date range

Modeling allocations in Cassandra

• Read all allocations for a customer – as fast as possible

• Find all of customers in a test/cell– reverse index

• Get count of customers in test/cell– count the entries in the reverse index

Denormalization - HOWTO

• No real world examples– ‘Normalization is for sissies’, Pat Helland

• Denormalize allocations per customer– Trivial with a schema-less database

Denormalized allocations

• normalized data

• denormalized (sparse) data

Implementing allocations

• As allocation for a customer has a handful of data points, they logically can be grouped together

• Avoided blobs, json or otherwise

• Using a standard column family, with composite columns

Composite columns

• Composite columns are sorted by each ‘token’ in name

• Allocation column naming convention– <testId>:<field>– 42:cell = 2– 42:enabled = Y– 47:cell = 0– 47:enabled = Y

Modeling AB metadata in cassandra

• Explored several models, including json blobs, spreading across multiple CFs, differing degrees of denormalization

• Reverse index to identify all tests for loading

Implementing metadata

• One CF, one row for all test’s data– Every data point is a column – no blobs

• Composite columns– type:id:field• Types = base info, cells, allocation plans• Id = cell number, allocation plan (gu)id• Field = type-specific

– Base info = test name, description, enabled– Cell’s name / description– Plan’s start/end dates, country to allocate to

Implementing indices

• Cassandra’s secondary indices vs. hand-built and maintained alternate indices

• Secondary indices work great on uniform data between rows

• But sparse column data not easy to index

Hand-built Indices, 1

• Reverse index– Test/cell (key) to custIds (columns) • Column value is timestamp

• Updating index when allocating a customer into test (double write)

Hand-built indices, 2

• Counter column family– Test/cell to count of customers in test columns– Mutate on allocating a customer into test

• Counters are not idempotent!• Mutates need to write to every node that

hosts that key

Index rebuilding

• To keep the index consistent, it needs to be rebuilt occasionally

• Even Oracle needs to have it’s indices rebuilt

Into the real world

Cassandra java clients

• Hector– github.com/rantav/hector

• Astyanax– Developed at Netflix (Eran Landau)– github.com/netflix

• Cassie (scala)– Developed at Twitter– https://github.com/twitter/cassie

Astyanax features

• Clean object model• Node discovery• Node quarantine• Request failover/retry• JMX Monitoring• Connection pooling• Future execution

Astyanax code example,1

Astyanax code example, 2

Astyanax code example, 3

Astyanax connection pools, 1

• Round Robin uses coordinator node

Astyanax connection pooling, 2

• Token aware knows where the data resides for point reads

Astyanax latency aware

• Samples response times from Cassandra nodes

• Favors faster responding nodes in pool• Use with token aware connection pooling

Allocation mutates

• AB allocations are immutable, so we need to prevent mutating

• Oracle - unique table constraint • Cassandra - read before write– data race!

Running cassandra

• Compactions happen– how Cassandra is maintained– Mutations are written to memory (Memtable)– Flushed to disk (SSTable) on triggering threshold– Eventually, Cassandra merges SSTables as data for

individual rows becomes scattered

Compactions, 2

• Latency spikes happen, especially on read-heavy systems– Everything can slow down– Throttling in newer Cassandra versions helps– Astyanax avoids this problem with latency

awareness

Tunings, 1

• Key and row caches– Left unbounded can consume JVM memory

needed for normal work– Latencies will spike as the JVM fights for free

memory– Off-heap row cache is better but still maintains

data structures on-heap

Tunings, 2

• mmap() as in-memory cache– When the Cassandra process is terminated, mmap

pages are returned to the free list• Row cache helps at startup

Tunings, 3

• Sizing memtable flushes for optimizing compactions– Easier when writes are uniformly distributed,

timewise – easier to reason about flush patterns– Best to optimize flushes based on memtable size,

not time

Tunings, 4

• Sharding – If a single row has disproportionately high

gets/mutates, the nodes holding it will become hot spots

– If a row grows too large, it can’t fit into memory

Takeaways

• Netflix is making all of our components distributed and fault tolerant as we grow domestically and internationally.

• Cassandra is a core piece of our cloud infrastructure.

• Netflix is open sourcing it’s cloud platform, including Cassandra support

終わり (The End)

• Q & A

@jasobrown jasedbrown@gmail.com

http://www.linkedin.com/in/jasedbrown

References

• Pat Helland, ‘Normalization Is for Sissies” http://blogs.msdn.com/b/pathelland/archive/2007/07/23/normalization-is-for-sissies.aspx