Pools, nodes, and load balancing methods
Domain 3 covers nodes, pool members and the algorithms that choose one. The exam probes the difference between member-level and node-level metrics, the state model behind a green pool, and how ratio and priority interact.
Objects and state
- A node is a server address; a member is a node plus a port inside a pool. A node marked down by a node-level monitor takes every member on it down.
- Availability comes from monitors (up, down, checking, unknown); Status is the administrative state (enabled/disabled).
- Disabling a member's session stops new connections; disabling its monitor marks it up without health checking.
Algorithms
- Round Robin, Ratio (static weights), Least Connections and Weighted Least Connections, Least Sessions, Observed, Predictive, Dynamic Ratio, Fastest, Hash, Random.
- Least Connections (member) compares per-member connection counts; (node) compares totals, so a node hosting several members attracts more traffic.
- Observed starts as round robin, then weights by a decaying connection count; Predictive projects the trend. Both settle over time rather than instantly.
- Dynamic Ratio needs a monitor that reports a metric (SNMP DCA or WMI); without one it falls back to static weights.
- Priority group activation keeps higher-priority members active and only adds the next group when the active count falls below the configured threshold.
Worked example
Pool members: A and B priority 1, C priority 2, D priority 3, with priority group activation set to "Less than 2 active members". Normally A and B serve. A fails; only one priority-1 member remains, which is fewer than 2, so C (priority 2) becomes active. D stays out until active priority-2 members drop below the threshold too. Within each active group, the load balancing method still decides distribution.
Traps
- Assuming a green pool means every member is monitored; a member with monitoring disabled stays available regardless of health.
- Using Dynamic Ratio without a metric-producing monitor.
- Expecting Observed to behave immediately; the first connections are round robin.
- Setting a ratio of zero and expecting the member to receive no connections at all.
- Confusing node statistics with member statistics when diagnosing uneven distribution.