Production Checklist

Deploying Riak KV to a realtime production environment from a development or testing environment can be a complex process. While the specifics of that process will always depend on your environment and practices, there are some basics for you to consider and a few questions that you will want to ask while making this transition.

We’ve compiled these considerations and questions into separate categories for you to look over.

System

Network

  • Are all systems using the same NTP servers to synchronize clocks?
  • Are you sure that your NTP clients’ configuration is monotonic (i.e. that your clocks will not roll back)?
  • Is DNS correctly configured for all systems’ production deployments?
  • Are connections correctly routed between all Riak nodes?
  • Are connections correctly set up in your load balancer?
  • Are your firewalls correctly configured?
  • Check that network latency and throughput are as expected for all of the following (we suggest using iperf to verify):
    • between nodes in the cluster
    • between the load balancer and all nodes in the cluster
    • between application servers and the load balancer
  • Do all Riak nodes appear in the load balancer’s rotation?
  • Is the load balancer configured to balance connections with roundrobin or a similarly random distribution scheme?

Riak KV

  • Check configuration files:
    • Does each machine have the correct name and IP settings in riak.conf (or in app.config if you’re using the older configuration files)?
    • Are all configurable settings identical across the cluster?
    • Have all of the settings in your configuration file(s) that were changed for debugging purposes been reverted back to production settings?
    • If you’re using multiple data backends, are all of your bucket types configured to use the correct backend?
    • If you are using Riak Security, have you checked off all items in the security checklist and turned on security?
    • If you’re using multiple data backends, do all machines’ config files agree on their configuration?
    • Do all nodes agree on the value of the [allow_mult][config basic] setting?
    • Do you have a sibling resolution strategy in place if allow_mult is set to true?
    • Have you carefully weighed the consistency trade-offs that must be made if allow_mult is set to false?
    • Are all of your apps replication properties configured correctly and uniformly across the cluster?
    • If you are using Riak Search, is it enabled on all nodes? If you are not, has it been disabled on all nodes?
    • If you are using strong consistency for some or all of your data:
      • Does your cluster consist of at least three nodes? If it does not, you will not be able to use this feature, and you are advised against enabling it.
      • If your cluster does consist of at least three nodes, has the strong consistency subsystem been [enabled][config strong consistency] on all nodes?
      • Is the target_n_val that is set on each node higher than any n_val that you intend to use for strongly consistent bucket types (or any bucket types for that matter)? The default is 4, which will likely need to be raised if you are using strong consistency.
    • Have all bucket types that you intend to use been created and successfully activated?
    • If you are using riak_control, is it enabled on the node(s) from which you intend to use it?
  • Check data mount points:
    • Is /var/lib/riak mounted?
    • Can you grow that disk later when it starts filling up?
    • Do all nodes have their own storage systems (i.e. no SANs), or do you have a plan in place for switching to that configuration later?

  • Are all Riak KV nodes up?

    • Run riak ping on all nodes. You should get pong as a response.
    • Run riak-admin wait-for-service riak_kv <node_name>@<IP> on each node. You should get riak_kv is up as a response.

    The <node_name>@<IP> string should come from your [configuration file(s)][configure reference].

  • Do all nodes agree on the ring state?

    • Run riak-admin ringready. You should get TRUE ALL nodes agree on the ring [list_of_nodes].
    • Run riak-admin member-status. All nodes should be valid (i.e. listed as Valid: 1), and all nodes should appear in the list
    • Run riak-admin ring-status. The ring should be ready (Ring Ready: true), there should be no unreachable nodes (All nodes are up and reachable), and there should be no pending changes to the ring (No pending changes).
    • Run riak-admin transfers. There should be no active transfers (No transfers active).

Operations

  • Does your monitoring system ensure that NTP is running?
  • Are you collecting time series data on the whole cluster?
    • System metrics
    • CPU load
    • Memory used
    • Network throughput
    • Disk space used/available
    • Disk input/output operations per second (IOPS)
    • Riak metrics (from the /stats HTTP endpoint or using riak-admin)
    • Latencies: GET and PUT (mean/median/95th/99th/100th)
    • Vnode stats: GETs, PUTs, GET totals, PUT totals
    • Node stats: GETs, PUTs, GET totals, PUT totals
    • Finite state machine (FSM) stats:
      • GET/PUT FSM objsize (99th and 100th percentile)
      • GET/PUT FSM times (mean/median/95th/99th/100th)
    • Protocol buffer connection stats
      • pbc_connects
      • pbc_active
      • pbc_connects_total
  • Are the following being graphed (at least the key metrics)?
    • Basic system status
    • Median and 95th and 99th percentile latencies (as these tend to be leading indicators of trouble)

Application and Load

  • Have you benchmarked your cluster with simulated load to confirm that your configuration will meet your performance needs?
  • Are the [develop client libraries] in use in your application up to date?
  • Do the client libraries that you’re using support the version of Riak KV that you’re deploying?

Confirming Configuration with Riaknostic

Recent versions of Riak KV ship with Riaknostic, a diagnostic utility that can be invoked by running riak-admin diag <check>, where check is one of the following:

  • disk
  • dumps
  • memory_use
  • nodes_connected
  • ring_membership
  • ring_preflists
  • ring_size
  • search
  • sysctl

Running riak-admin diag with no additional arguments will run all checks and report the findings. This is a good way of verifying that you’ve gotten at least some of the configurations mentioned above correct, that all nodes in your cluster are up, and that nothing is grossly misconfigured. Any warnings produced by riak-admin diag should be addressed before going to production.

Troubleshooting and Support

  • Does your team, including developing and operations, know how to open support requests with Riak?
  • Is your team familiar with Riak Support’s Service-Level Agreement (SLA) levels?
    • Normal and Low are for issues not immediately impacting production systems
    • High is for problems that impact production or soon-to-be-production systems, but where stability is not currently compromised
    • Urgent is for problems causing production outages or for those issues that are likely to turn into production outages very soon. On-call engineers respond to urgent requests within 30 minutes, 24 / 7.
  • Does your team know how to gather riak-debug results from the whole cluster when opening tickets? If not, that process goes something like this:
    • SSH into each machine, run riak-debug, and grab the resultant .tar.gz file
    • Attach all debug tarballs from the whole cluster each time you open a new High- or Urgent-priority ticket

The Final Step: Taking it to Production

Once you’ve been running in production for a month or so, look back at the metrics gathered above. Based on the numbers you’re seeing so far, configure alerting thresholds on your latencies, disk consumption, and memory. These are the places most likely to give you advance warning of trouble.

When you go to increase capacity down the line, having historic metrics will give you very clear indicators of having resolved scaling problems, as well as metrics for understanding what to upgrade and when.