Collection Dimension | Collection methods | Representative Metric |
CPU resources | Collect the CPU utilization of Proxy processes in real time. | cpu_util |
Request statistics | Collect command execution counts and type distribution through Proxy. | proxy_commands,cmd_key_count,cmd_mget,cmd_err,cmd_big_value |
Network traffic | Collect inbound and outbound network traffic and bandwidth utilization. | in_flow,out_flow,in_bandwidth_util,out_bandwidth_util |
Traffic throttling | Record the number of times bandwidth throttling is triggered. | in_flow_limit,out_flow_limit |
Connection management | Collect the number and utilization of client TCP connections. | connections,connections_util,connections_max_util |
Execution latency | Collect command execution latency distribution between Proxy and Redis Server. | latency_avg,latency_max,latency_read,latency_write,latency_other |
Collection methods | Description | Representative Metric |
Real-time status collection | Directly read the status information of the data node at runtime to obtain the instantaneous value at the current moment. | cpu_util (CPU utilization), connections (number of connections), mem_used (memory usage), mem_util (memory utilization), keys (total number of keys), expires (number of keys with expiration time set) |
Periodic increment calculation | In each collection cycle, calculate the change in the metric value relative to the previous cycle, reflecting the incremental change within the collection cycle. | expired (number of expired keys), evicted (number of evicted keys), cmd_slow (number of slow queries), cmd_hits (number of read request hits), cmd_miss (number of read request misses) |
Rate calculation | Divide the result of the incremental calculation by the collection interval to obtain the per-second rate, reflecting the processing capability per unit time. | commands (total request QPS), cmd_read (read request QPS), cmd_write (write request QPS), cmd_other (other request QPS) |
Combined derivative calculation | Generate derived metrics based on combined calculations of multiple basic metrics, reflecting higher-level performance characteristics. | cmd_hits_ratio (read request hit ratio = hits / (hits + misses)) |
Aggregation Algorithm | Scenarios | Representative Metric | Calculation Method |
SUM | Countable metrics that can be summed, where the combined values of all nodes represent the total instance count. | commands (total request QPS), cmd_read (read request QPS), cmd_write (write request QPS), cmd_slow (slow query count), cmd_hits (hit count), cmd_miss (miss count) | Sum the metric values of all nodes involved in aggregation. |
AVG | Ratio metrics that need to reflect the overall average level of an instance | cpu_util (CPU utilization) | Calculate the arithmetic mean of metric values across all Redis nodes, including master and replica nodes. |
MAX (maximum value) | Metrics that need to reflect the bottleneck node status within an instance | cpu_max_util (maximum node CPU utilization), mem_max_util (maximum node memory utilization), connections_max_util (maximum node connection utilization) | Take the maximum metric value among all nodes involved in aggregation. |
Weighted average | Ratio metrics whose global characteristics cannot be accurately reflected by simple averaging need to be weighted based on the actual data volume of each node. | cmd_hits_ratio (read request hit ratio) | Total hits of all nodes / (total hits of all nodes + total misses of all nodes), rather than the simple average of the hit ratio of each node. |
LAST (latest value) | Take the value from the latest collection. | Some status metrics | Take the value from the most recent collection cycle. |
cmd_hits_ratio) is calculated using a weighted average algorithm, which more accurately reflects the overall cache hit status of the instance compared with a simple average. For example, when access volumes vary significantly across nodes, a simple average may be skewed by low-traffic nodes, whereas a weighted average is calculated based on the actual total number of hits and misses, producing more accurate data. When the instance has no access requests, this value is null.Metric | Metric Name | Description |
Total request tilt rate | qps_slope_util | Reflects the balance of request volume across shards. |
Memory utilization tilt rate | mem_util_slope_util | Reflects the balance of memory usage across shards. |
Scenario | Impact | Description |
Instance initialization | Brief fluctuations in some metrics | Incremental calculation metrics (such as QPS and slow query quantity) require data from consecutive collection cycles to output stable values, and may briefly show zero or abnormal values during the initial startup phase. |
Primary-secondary switch | Brief data interruption | During a failover, there may be a brief data gap between when the original node stops reporting and when the new node starts reporting. After the switchover is complete, monitoring data is automatically restored. |
Scale-out or scale-in | Node ID change | Scale-out and scale-in operations cause node ID changes. If you query monitoring data by node through the API, you need to obtain node information again after the operation is complete. For details, see DescribeInstanceNodeInfo. |
Enabling or disabling read-only replica | Instance-level metric baseline change | After enabling or disabling the read-only replica feature, the scope of nodes involved in instance-level aggregate calculation may change, and the value baselines of some metrics may be adjusted accordingly. |
피드백