Distributed Erlang Systems In Operation
7 likes2,322 views
Distributed Erlang systems aim to be decentralized, distributed, homogeneous, and fault tolerant. Nodes use only local data and there is no global state or reliance on physical time. Cluster membership is handled through a configuration file or gossip protocol. Load balancing uses techniques like consistent hashing. Liveness is checked through Erlang tools and custom mechanisms. Soft state through gossip protocols can provide an alternative to global state. Shipping code embedded in a runtime and upgrading through reboots helps manage distributed systems.
![Load Balancing and
Resource Allocation
• Static assignment
• Round-robin/Random
• Static hashing: Nodes[hash(Item) mod
length(Nodes)]
• Dynamo/Riak/Cassandra/Voldemort:
Consistent Hashing](https://image.slidesharecdn.com/argef2010-100325150703-phpapp02/85/Distributed-Erlang-Systems-In-Operation-7-320.jpg)
![Questions?
“You know you have [a distributed system]
when the crash of a computer you’ve never
heard of stops you from getting any work
done”
-Leslie Lamport](https://image.slidesharecdn.com/argef2010-100325150703-phpapp02/85/Distributed-Erlang-Systems-In-Operation-15-320.jpg)
Distributed Erlang Systems In Operation
- 1. Distributed Erlang Systems In Operation Andy Gross <[email protected]>, @argv0 VP, Engineering Basho Technologies Erlang Factory SF 2010
- 2. Architectural Goals • Decentralized (no masters). • Distributed (asynchronous, nodes use only local data). • Homogeneous (all nodes can do anything). • Fault tolerant (emergent goal). • Observable
- 3. Anti-Goals • Global state: • pg2/hot data in mnesia • globally registered names • Distributed transactions • Reliance on physical time
- 4. Compromise your Goals • Decentralized (no masters). • Distributed (nodes use only local data). • Homogeneous (all nodes can do anything). • No distributed transactions/global state. • No reliance on physical time.
- 5. Systems Design • Cluster Membership • Load balancing/naming/resource allocation • Liveness checking • Soft Global State
- 6. Cluster Membership • Option 1: Use a configuration file: • Requires out-of-band sync of configuration file across machines. • Not “elastic” enough for some use-cases. • Option II: Contact a seed node to join and use gossip protocol to propagate state.
- 7. Load Balancing and Resource Allocation • Static assignment • Round-robin/Random • Static hashing: Nodes[hash(Item) mod length(Nodes)] • Dynamo/Riak/Cassandra/Voldemort: Consistent Hashing
- 8. Liveness Checking • nodes() and net_adm:ping() operations can be too low-level. • Sometimes you’d like to divert traffic from a node at the application level while keeping distributed Erlang up. • Use net_kernel:monitor_nodes() and an app-level mechanism for liveness.
- 9. Soft State/Gossip Protocols • An eventually-consistent alternative to global state. • Nodes make changes, gossip to another node. • Nodes receive changes, merge with local state, gossip to another node. • Requires up-front thought about data structures, dealing with slightly-stale data.
- 10. Running Your System • Shipping code • Upgrading code • Debugging your own systems • Living with other people’s systems
- 11. Shipping Code • Don’t rely on working Erlang on end-user machines (many Linux distros are broken or out of date). • Ship code with an embedded runtime and libraries. • Put version/build info in code.
- 12. Upgrading Code • Hot code loading for small, emergency fixes. • For new releases, reboot the node. • Why not .appups? • Systems I’ve worked on have changed/ evolved too fast. • A reboot is a good test of resiliency.
- 13. Debugging Running Systems • Remote Erlang shells are awesome, except when distributed Erlang dies (it happens). • run_erl (or even screen(1)) give you a backdoor for when -remsh fails. • rebar (http://hg.basho.com/rebar) makes this easy. • What if you don’t have access to the box?
- 14. OPS - Other People’s Systems • Your Erlang, Enterprise firewalls. • Erlang shell is powerful, but scary. • Provide a debugging module. • Get data out via HTTP/SMTP/SNMP • Use disk_log/report_browser.
- 15. Questions? “You know you have [a distributed system] when the crash of a computer you’ve never heard of stops you from getting any work done” -Leslie Lamport
- 16. Resources • unsplit: http://github.com/uwiger/unsplit • gen_leader: http://github.com/KirinDave/gen_leader_revival • Dynamo: http://www.allthingsdistributed.com/2007/10/ amazons_dynamo.html • Hans Svensson: Distributed Erlang Application Pitfalls and Recipes: http://www.erlang.org/workshop/2007/proceedings/ 06svenss.ppt • Consistent Hashing and Random Trees: Distributed Caching Protocols for relieving Hot Spots on the World Wide Web: http://bit.ly/LewinConsistentHashing