Adding a distributed table with remote agents

Please read the article about distributed tables for general overview of distributed tables. Here we focus on using a distributed table as a basis for creating a cluster of Manticore instances.

Here we have split the data over 4 servers, each serving one of the shards. If one of the servers fails, our distributed table will still work, but we would miss the results from the failed shard.

  • ini

ini

  1. table mydist {
  2.           type  = distributed
  3.           agent = box1:9312:shard1
  4.           agent = box2:9312:shard2
  5.           agent = box3:9312:shard3
  6.           agent = box4:9312:shard4
  7. }

Now that we’ve added mirrors, each shard is found on 2 servers. By default, the master (the searchd instance with the distributed table) will pick randomly one of the mirrors.

The mode used for picking mirrors can be set with ha_strategy. In addition to the default random mode there’s also ha_strategy = roundrobin.

More interesting strategies are those based on latency-weighted probabilities. noerrors and nodeads: not only those take out mirrors with issues, but also monitor the response times and do balancing. If a mirror responds slower (for example due to some operations running on it), it will receive less requests. When the mirror recovers and provides better times, it will get more requests.

  • ini

ini

  1. table mydist {
  2.           type  = distributed
  3.           agent = box1:9312|box5:9312:shard1
  4.           agent = box2:9312:|box6:9312:shard2
  5.           agent = box3:9312:|box7:9312:shard3
  6.           agent = box4:9312:|box8:9312:shard4
  7. }

Mirroring

Agent mirrors can be used interchangeably when processing a search query. Manticore instance (can be multiple) hosting the distributed table where the mirrored agents are defined keeps track of mirror status (alive or dead) and response times, and does automatic failover and load balancing based on that.

Agent mirrors

  1. agent = node1|node2|node3:9312:shard2

The above example declares that ‘node1:9312’, ‘node2:9312’, and ‘node3:9312’ all have a table called shard2, and can be used as interchangeable mirrors. If any single of those servers go down, the queries will be distributed between the other two. When it gets back up, master will detect that and begin routing queries to all three nodes again.

Mirror may also include individual table list, as:

  1. agent = node1:9312:node1shard2|node2:9312:node2shard2

This works essentially the same as the previous example, but different table names will be used when querying different severs: node1shard2 when querying node1:9312, and node2shard when querying node2:9312.

By default, all queries are routed to the best of the mirrors. The best one is picked based on the recent statistics, as controlled by the ha_period_karma config directive. Master stores a number of metrics (total query count, error count, response time, etc) recently observed for every agent. It groups those by time spans, and karma is that time span length. The best agent mirror is then determined dynamically based on the last 2 such time spans. Specific algorithm that will be used to pick a mirror can be configured ha_strategy directive.

The karma period is in seconds and defaults to 60 seconds. Master stores up to 15 karma spans with per-agent statistics for instrumentation purposes (see SHOW AGENT STATUS statement). However, only the last 2 spans out of those are ever used for HA/LB logic.

When there are no queries, master sends a regular ping command every ha_ping_interval milliseconds in order to have some statistics and at least check, whether the remote host is still alive. ha_ping_interval defaults to 1000 msec. Setting it to 0 disables pings and statistics will only be accumulated based on actual queries.

Example:

  1. # sharding table over 4 servers total
  2. # in just 2 shards but with 2 failover mirrors for each shard
  3. # node1, node2 carry shard1 as local
  4. # node3, node4 carry shard2 as local
  6. # config on node1, node2
  7. agent = node3:9312|node4:9312:shard2
  9. # config on node3, node4
  10. agent = node1:9312|node2:9312:shard1

Load balancing

Load balancing is turned on by default for any distributed table using mirroring. By default queries are distributed randomly among the mirrors. To change this behaviour you can use ha_strategy.

ha_strategy

  1. ha_strategy = {random|nodeads|noerrors|roundrobin}

Agent mirror selection strategy for load balancing. Optional, default is random.

The strategy used for mirror selection, or in other words, choosing a specific agent mirror in a distributed table. Essentially, this directive controls how exactly master does the load balancing between the configured mirror agent nodes. The following strategies are implemented:

Simple random balancing

The default balancing mode. Simple linear random distribution among the mirrors. That is, equal selection probability are assigned to every mirror. Kind of similar to round-robin (RR), but unlike RR, does not impose a strict selection order.

  • Example

Example

  1. ha_strategy = random

Adaptive randomized balancing

The default simple random strategy does not take mirror status, error rate and, most importantly, actual response latencies into account. So to accommodate for heterogeneous clusters and/or temporary spikes in agent node load, we have a group of balancing strategies that dynamically adjusts the probabilities based on the actual query latencies observed by the master.

The adaptive strategies based on latency-weighted probabilities basically work as follows:

  • latency stats are accumulated in blocks of ha_period_karma seconds;
  • once per karma period latency-weighted probabilities get recomputed;
  • once per request (including ping requests) “dead or alive” flag is adjusted.

Currently, we begin with equal probabilities (or percentages, for brevity), and on every step, scale them by the inverse of the latencies observed during the last “karma” period, and then renormalize them. For example, if during the first 60 seconds after the master startup 4 mirrors had latencies of 10, 5, 30, and 3 msec/query respectively, the first adjustment step would go as follow:

  • initial percentages: 0.25, 0.25, 0.25, 0.25;
  • observed latencies: 10 ms, 5 ms, 30 ms, 3 ms;
  • inverse latencies: 0.1, 0.2, 0.0333, 0.333;
  • scaled percentages: 0.025, 0.05, 0.008333, 0.0833;
  • renormalized percentages: 0.15, 0.30, 0.05, 0.50.

Meaning that the 1st mirror would have a 15% chance of being chosen during the next karma period, the 2nd one a 30% chance, the 3rd one (slowest at 30 ms) only a 5% chance, and the 4th and the fastest one (at 3 ms) a 50% chance. Then, after that period, the second adjustment step would update those chances again, and so on.

The rationale here is, once the observed latencies stabilize, the latency weighted probabilities stabilize as well. So all these adjustment iterations are supposed to converge at a point where the average latencies are (roughly) equal over all mirrors.

nodeads

Latency-weighted probabilities, but dead mirrors are excluded from the selection. “Dead” mirror is defined as a mirror that resulted in multiple hard errors (eg. network failure, or no answer, etc) in a row.

  • Example

Example

  1. ha_strategy = nodeads

noerrors

Latency-weighted probabilities, but mirrors with worse errors/success ratio are excluded from the selection.

  • Example

Example

  1. ha_strategy = noerrors

Round-robin balancing

Simple round-robin selection, that is, selecting the 1st mirror in the list, then the 2nd one, then the 3rd one, etc, and then repeating the process once the last mirror in the list is reached. Unlike with the randomized strategies, RR imposes a strict querying order (1, 2, 3, …, N-1, N, 1, 2, 3, … and so on) and guarantees that no two subsequent queries will be sent to the same mirror.

  • Example

Example

  1. ha_strategy = roundrobin

Instance-wide options

ha_period_karma

  1. ha_period_karma = 2m

ha_period_karma defines agent mirror statistics window size, in seconds (or time suffixed). Optional, default is 60.

For a distributed table with agent mirrors in it, server tracks several different per-mirror counters. These counters are then used for failover and balancing. (Server picks the best mirror to use based on the counters.) Counters are accumulated in blocks of ha_period_karma seconds.

After beginning a new block, master may still use the accumulated values from the previous one, until the new one is half full. Thus, any previous history stops affecting the mirror choice after 1.5 times ha_period_karma seconds at most.

Despite that at most 2 blocks are used for mirror selection, up to 15 last blocks are actually stored, for instrumentation purposes. They can be inspected using SHOW AGENT STATUS statement.

ha_ping_interval

  1. ha_ping_interval = 3s

ha_ping_interval defines interval between agent mirror pings, in milliseconds (or time suffixed). Optional, default is 1000.

For a distributed table with agent mirrors in it, server sends all mirrors a ping command during the idle periods. This is to track the current agent status (alive or dead, network roundtrip, etc). The interval between such pings is defined by this directive.

To disable pings, set ha_ping_interval to 0.