Labels: mode — subset (the sample scanned a subset of the files, sized by tablet_pre_split_data_tier_scan_byte_limit), under_limit (the input was within the limit, so every file was scanned), disabled (the limit is 0, so every file was scanned), all_selected (the input exceeded the limit, but taking whole files, at least one per partition and at least tablet_pre_split_data_tier_min_scan_files in all, still selected every file), partition_from_file_data (the input exceeded the limit, but a partition column is read from the file data rather than from the path, or path and literal partition columns are mixed, so every file was scanned: a subset could miss whole partitions), path_not_expressible (INSERT INTO ... SELECT FROM FILES() only: a selected file's path contains a character FILES cannot take as an exact path — ,, *, ?, [, {, \, a : in the path after the host, or leading or trailing whitespace — so the sample scanned every file through the statement's own path).
Description: Total data-tier samples of Broker Load and INSERT INTO ... SELECT FROM FILES() loads, by how the scanned files were chosen.
Description: Share of the input bytes that each data-tier sample of a Broker Load or INSERT INTO ... SELECT FROM FILES() load scanned, as a percentage rounded down to an integer, so 100 means every file was scanned.
Description: Cumulative number of task exceptions caught and swallowed by ThreadPool worker threads across the BE process. Increments only when enable_threadpool_catch_task_exception is true. When that item is false (default), this metric stays unchanged because there is no enclosing catch clause. Use it to alert on swallowed failures while catch mode is enabled; pool name and exception detail remain in the BE ERROR logs.
Description: Total number of retries that FE Thrift RPC calls made after their connection broke. Before each retry, the FE waits a random interval with exponential backoff (100 to 200 ms before the first retry, with both bounds doubling for each further retry until the wait reaches 500 to 1000 ms), then reopens the connection and resends the request. Calls that time out are not retried and are not counted. The number of retries per call is bounded by thrift_rpc_retry_times. A rising value means the FE keeps losing its Thrift connections to other nodes, for example because the Thrift worker pool of the Leader FE is saturated and closing new connections; read it together with thrift_server_rejected_connections_total on that node.
Description: Milliseconds since the FE Thrift accept loop last returned a connection. A wedged acceptor makes this climb, but so does an FE with no Thrift traffic, so read it together with the connection arrival rate rather than alerting on it alone. 0 also covers the case where no acceptor is running at all, before the Thrift server starts and after it stops, so a threshold alert on this metric cannot tell a dead Thrift service from a healthy one.
Description: Total number of connections that the FE Thrift server closed without serving because they had waited in the pending queue longer than thrift_server_queue_timeout_ms. That check is disabled by default, so this counter stays at 0 until an operator arms the timeout. A rising value means the FE was working through a backlog of connections whose callers had almost certainly given up already; read it together with thrift_server_queue_wait_ms to see how far behind the queue had fallen.
Description: Quantiles of how long connections waited in the FE Thrift server pending queue before a worker thread picked them up. Sampled for every dequeued connection, including the ones then dropped as expired, so the distribution is not truncated at thrift_server_queue_timeout_ms. This is the leading indicator of Thrift saturation: it climbs while the worker pool is still keeping up, well before thrift_server_rejected_connections_total starts to move.
Description: Total number of connections the FE Thrift server closed immediately because its worker pool was saturated. Every rejection is counted, including those whose log warning was suppressed by rate limiting. A rising value means clients are being turned away; read it with the thread_pool metrics for thrift-server-pool, where the rate of increase indicates how far worker capacity is short.
Description: Shared-data only. Number of auto vacuum rounds on the leader FE in which a Vacuum request for a partition could not be sent or returned an error.
Description: Shared-data only. Number of auto vacuum rounds on the leader FE in which every Vacuum request sent to the CNs for a partition returned success. Rounds that send no request are not counted.