Billions of Hits: Scaling Twitter John Adams Twitter Operations
John Adams @netik • Early Twitter employee (mid-2008) • Lead engineer: Outward Facing Services (Apache, Unicorn, SMTP), Auth, Security • Keynote Speaker: O’Reilly Velocity 2009 • O’Reilly Web 2.0 Speaker (2008, 2010) • Previous companies: Inktomi, Apple, c|net • Working on Web Operations book with John Alspaw (flickr, etsy), out in June
752% 2008 Growth source: comscore.com - (based only on www traffic, not API)
1358% 2009 Growth source: comscore.com - (based only on www traffic, not API)
12th most popular source: alexa.com
55M Tweets per day (640 TPS/sec, 1000 TPS/sec peak) source: twitter.com internal
600M Searches/Day source: twitter.com internal
25% Web API 75%
Operations • What do we do? • Site Availability • Capacity Planning (metrics-driven) • Configuration Management • Security • Much more than basic Sysadmin
What have we done? • Improved response time, reduced latency • Less errors during deploys (Unicorn!) • Faster performance • Lower MTTD (Mean time to Detect) • Lower MTTR (Mean time to Recovery)
Operations Mantra Move to Find Take Next Weakest Corrective Weakest Point Action Point Metrics + Logs + Science = Process Repeatability Analysis
Make an attack plan. Symptom Bottleneck Vector Solution HTTP Bandwidth Network Latency Servers++ Timeline Update Better Delay Database Delay algorithm Status Flock Growth Database Delays Cassandra Updates Algorithm Latency Algorithms
Finding Weakness • Metrics + Graphs • Individual metrics are irrelevant • We aggregate metrics to find knowledge • Logs • SCIENCE!
Monitoring • Twitter graphs and reports critical metrics in as near real time as possible • If you build tools against our API, you should too. • RRD, other Time-Series DB solutions • Ganglia + custom gmetric scripts • dev.twitter.com - API availability
Analyze • Turn data into information • Where is the code base going? • Are things worse than they were? • Understand the impact of the last software deploy • Run check scripts during and after deploys • Capacity Planning, not Fire Fighting!
Data Analysis • Instrumenting the world pays off. • “Data analysis, visualization, and other techniques for seeing patterns in data are going to be an increasingly valuable skill set. Employers take notice!” “Web Squared: Web 2.0 Five Years On”, Tim O’Reilly, Web 2.0 Summit, 2009
Forecasting Curve-fitting for capacity planning (R, fityk, Mathematica, CurveFit) unsigned int (32 bit) Twitpocolypse status_id signed int (32 bit) Twitpocolypse r2=0.99
What’s a Robot ? • Actual error in the Rails stack (HTTP 500) • Uncaught Exception • Code problem, or failure / nil result • Increases our exception count • Shows up in Reports
What’s a Whale ? • HTTP Error 502, 503 • Twitter has a hard and fast five second timeout • We’d rather fail fast than block on requests • We also kill long-running queries (mkill) • Timeout
Whale Watcher • Simple shell script, • MASSIVE WIN by @ronpepsi • Whale = HTTP 503 (timeout) • Robot = HTTP 500 (error) • Examines last 60 seconds of aggregated daemon / www logs • “Whales per Second” > Wthreshold • Thar be whales! Call in ops.
Deploy Watcher Sample window: 300.0 seconds First start time: Mon Apr 5 15:30:00 2010 (Mon Apr 5 08:30:00 PDT 2010) Second start time: Tue Apr 6 02:09:40 2010 (Mon Apr 5 19:09:40 PDT 2010) PRODUCTION APACHE: ALL OK PRODUCTION OTHER: ALL OK WEB0049 CANARY APACHE: ALL OK WEB0049 CANARY BACKEND SERVICES: ALL OK DAEMON0031 CANARY BACKEND SERVICES: ALL OK DAEMON0031 CANARY OTHER: ALL OK
Feature “Darkmode” • Specific site controls to enable and disable computationally or IO-Heavy site function • The “Emergency Stop” button • Changes logged and reported to all teams • Around 60 switches we can throw • Static / Read-only mode
Servers • Co-located, dedicated machines at NTT America • No clouds; Only for monitoring, not serving • Need raw processing power, latency too high in existing cloud offerings • Frees us to deal with real, intellectual, computer science problems. • Moving to our own data center soon
unicorn • A single socket Rails application Server (Rack) • Zero Downtime Deploys (!) • Controlled, shuffled transfer to new code • Less memory, 30% less CPU • Shift from mod_proxy_balancer to mod_proxy_pass • HAProxy, Ngnix wasn’t any better. really.
Rails • Mostly only for front-end. • Back end mostly Scala and pure ruby • Not to blame for our issues. Analysis found: • Caching + Cache invalidation problems • Bad queries generated by ActiveRecord, resulting in slow queries against the db • Queue Latency • Replication Lag
memcached • memcached isn’t perfect. • Memcached SEGVs hurt us early on. • Evictions make the cache unreliable for important configuration data (loss of darkmode flags, for example) • Network Memory Bus isn’t infinite • Segmented into pools for better performance
Loony • Central machine database (MySQL) • Python, Django, Paraminko SSH • Paraminko - Twitter OSS (@robey) • Ties into LDAP groups • When data center sends us email, machine definitions built in real-time
Murder • @lg rocks! • Bittorrent based replication for deploys • ~30-60 seconds to update >1k machines • P2P - Legal, valid, Awesome.
Kestrel • @robey • Works like memcache (same protocol) • SET = enqueue | GET = dequeue • No strict ordering of jobs • No shared state between servers • Written in Scala.
Asynchronous Requests • Inbound traffic consumes a unicorn worker • Outbound traffic consumes a unicorn worker • The request pipeline should not be used to handle 3rd party communications or back-end work. • Reroute traffic to daemons
Daemons • Daemons touch every tweet • Many different daemon types at Twitter • Old way: One daemon per type (Rails) • New way: Fewer Daemons (Pure Ruby) • Daemon Slayer - A Multi Daemon that could do many different jobs, all at once.
Disk is the new Tape. • Social Networking application profile has many O(ny) operations. • Page requests have to happen in < 500mS or users start to notice. Goal: 250-300mS • Web 2.0 isn’t possible without lots of RAM • SSDs? What to do?
Caching • We’re the real-time web, but lots of caching opportunity. You should cache what you get from us. • Most caching strategies rely on long TTLs (>60 s) • Separate memcache pools for different data types to prevent eviction • Optimize Ruby Gem to libmemcached + FNV Hash instead of Ruby + MD5 • Twitter now largest contributor to libmemcached
MySQL • Sharding large volumes of data is hard • Replication delay and cache eviction produce inconsistent results to the end user. • Locks create resource contention for popular data
MySQL Challenges • Replication Delay • Single threaded. Slow. • Social Networking not good for RDBMS • N x N relationships and social graph / tree traversal • Disk issues (FS Choice, noatime, scheduling algorithm)
Relational Databases not a Panacea • Good for: • Users, Relational Data, Transactions • Bad: • Queues. Polling operations. Social Graph. • You don’t need ACID for everything.
Database Replication • Major issues around users and statuses tables • Multiple functional masters (FRP, FWP) • Make sure your code reads and writes to the write DBs. Reading from master = slow death • Monitor the DB. Find slow / poorly designed queries • Kill long running queries before they kill you (mkill)
Flock • Flock Scalable Social Graph Store • Sharding via Gizzard • Gizzard MySQL backend (many.) • 13 billion edges, 100K reads/second • Mysql Mysql Mysql Open Source!
Cassandra • Originally written by Facebook • Distributed Data Store • @rk’s changes to Cassandra Open Sourced • Currently double-writing into it • Transitioning to 100% soon.
Lessons Learned • Instrument everything. Start graphing early. • Cache as much as possible • Start working on scaling early. • Don’t rely on memcache, and don’t rely on the database • Don’t use mongrel. Use Unicorn.