In an attempt to summarize rampconf 2013, we discuss the awesome devops concepts used by leading software firms:
Second class users
Reddit uses this awesome strategy of treating non-logged in users as second class users hitting the CDN rather than the servers.
Consistent hashing as disaster recovery
Netflix has these large clusters of cassandra servers. The data is replicated in stripes in various datacenters across various servers
Classify functions as security groups
LOG LOG LOG but do it right, adding context to the logs
Bottom-up Monitoring
In scenarios of realtime monitoring the top-down approach fails as a green server might also inculcate operational bugs. Solution is to add more context to an alert when something breaks. Top down monitoring only resolves to go/nogo, which doesnt suffice for most of the scenarios.
Preallocation strategy for OOMs for memory bound services
In a traffic to memory application types, you live in a constant fear that your traffic is going to increase and you eventually run out of memory. You preallocate memory as a reserve a and later release it when needed.
This strategy was used by indextools to adapt to quickly varying traffic patterns.
Service Oriented Architecture
Essentially
Neil J Gunther - Universal Law of Scalability
This aust
Tools like Bees with Machine Guns does a DDos attack on your own system to measure the performance
Similarly we have KillerBees from softonic
Metrics Are really important. Collect Application Metrics with a SaaS based tool called NewRelic
Factors: Contention and Coherency Delay
Share Nothing Architecture
Sometimes it makes sense to
1 > 2 > 3
Related to development cycles. Iteration cycles from 1 to 2 is more difficult than iteratioon cycles 2 > 3. Start with queues, appservers which can be more than 1 at any time of problem.
Event or Thread Based
Paxos Paper - libpaxos (alternatively raft), statsis- google code
riak-core. It gives the following
Quick Check again Unit Testing (Domain Related Test Cases )
e.g. a function takes an inpu
Pulse - Spoof the scheduler - Traces messages between processes and generates a pathological set of scenarios
Second class users
Reddit uses this awesome strategy of treating non-logged in users as second class users hitting the CDN rather than the servers.
Consistent hashing as disaster recovery
Netflix has these large clusters of cassandra servers. The data is replicated in stripes in various datacenters across various servers
Classify functions as security groups
LOG LOG LOG but do it right, adding context to the logs
Bottom-up Monitoring
In scenarios of realtime monitoring the top-down approach fails as a green server might also inculcate operational bugs. Solution is to add more context to an alert when something breaks. Top down monitoring only resolves to go/nogo, which doesnt suffice for most of the scenarios.
Preallocation strategy for OOMs for memory bound services
In a traffic to memory application types, you live in a constant fear that your traffic is going to increase and you eventually run out of memory. You preallocate memory as a reserve a and later release it when needed.
This strategy was used by indextools to adapt to quickly varying traffic patterns.
Service Oriented Architecture
Essentially
Neil J Gunther - Universal Law of Scalability
This aust
Tools like Bees with Machine Guns does a DDos attack on your own system to measure the performance
Similarly we have KillerBees from softonic
Metrics Are really important. Collect Application Metrics with a SaaS based tool called NewRelic
Factors: Contention and Coherency Delay
Share Nothing Architecture
Sometimes it makes sense to
1 > 2 > 3
Related to development cycles. Iteration cycles from 1 to 2 is more difficult than iteratioon cycles 2 > 3. Start with queues, appservers which can be more than 1 at any time of problem.
Event or Thread Based
Paxos Paper - libpaxos (alternatively raft), statsis- google code
riak-core. It gives the following
Quick Check again Unit Testing (Domain Related Test Cases )
e.g. a function takes an inpu
Pulse - Spoof the scheduler - Traces messages between processes and generates a pathological set of scenarios
No comments:
Post a Comment