Metrics Reference

Last update: October 13 2026Open in Markdown

This instruction describes the product metrics collected by the service.

The list of products that support metrics will be expanded.

Dedicated Servers

The dedicated servers namespace is dedicated.

Monitoring is supported by servers with a KVM console, which is available for every dedicated server, except servers of the Chipcore product line and servers in a dedicated rack — for details, see the Servers in a dedicated rack page on selectel.ru.

The set of IPMI metrics may differ depending on the server model.

Each metric includes a timestamp — the metric collection time. You can convert the value using online services, for example, on the unixtimestamp.com website.

Metric name Type, unit of measurement	Description	Labels
ipmi_bmc_info Gauge, UNIT	BMC (Baseboard Management Controller) metadata — firmware, manufacturer, and more. The metric value is always 1	datacenter — name of the data center where the server is located; location_id — pool where the server is located; resource_id — server UUID; bmc_url — BMC address; firmware_revision — BMC firmware version; manufacturer_id — manufacturer (vendor) ID; system_firmware_version — system firmware version (BIOS, UEFI)
up Gauge, bool	Check metric source availability: 1 — success; 0 — failure	datacenter — name of the data center where the server is located; location_id — pool where the server is located; resource_id — server UUID
ipmi_up Gauge, bool	Check IPMI availability: 1 — success; 0 — failure	datacenter — name of the data center where the server is located; collector — name of the exporter's internal collector responsible for gathering a specific group of metrics; location_id — pool where the server is located; resource_id — server UUID
ipmi_scrape_duration_seconds Gauge, seconds	Time taken to collect IPMI exporter metrics	datacenter — name of the data center where the server is located; location_id — pool where the server is located; resource_id — server UUID
ipmi_sel_events_count_by_state Gauge, count	Number of events from the SEL (System Event Log) journal, grouped by event severity level	datacenter — name of the data center where the server is located; location_id — pool where the server is located; resource_id — server UUID; state — event severity level
ipmi_chassis_cooling_fault_state Gauge, bool	Server cooling system state: 1 — healthy; 0 — failure	datacenter — name of the data center where the server is located; location_id — pool where the server is located; resource_id — server UUID
ipmi_chassis_drive_fault_state Gauge, bool	Drive state: 1 — healthy; 0 — failure	datacenter — name of the data center where the server is located; location_id — pool where the server is located; resource_id — server UUID
ipmi_chassis_power_state Gauge, bool	Server power state: 1 — powered on; 0 — powered off	datacenter — name of the data center where the server is located; location_id — pool where the server is located; resource_id — server UUID
ipmi_fan_speed_state Gauge, enum	Fan state: 0 — healthy; 1 — deviation; 2 — failure	datacenter — name of the data center where the server is located; location_id — pool where the server is located; resource_id — server UUID; id — sensor identifier; name — sensor name
ipmi_power_state Gauge, enum	Server component power state: 0 — healthy; 1 — deviation; 2 — failure	datacenter — name of the data center where the server is located; location_id — pool where the server is located; resource_id — server UUID; id — sensor ID; name — sensor name
ipmi_temperature_state Gauge, enum	Server component temperature state: 0 — healthy; 1 — deviation; 2 — failure	datacenter — name of the data center where the server is located; location_id — pool where the server is located; resource_id — server UUID; id — sensor ID; name — sensor name
ipmi_voltage_state Gauge, enum	Voltage state: 0 — healthy; 1 — deviation; 2 — failure	datacenter — name of the data center where the server is located; location_id — pool where the server is located; resource_id — server UUID; id — sensor ID; name — sensor name
ipmi_sensor_state Gauge, enum	Server component state: 0 — healthy; 1 — deviation; 2 — failure	datacenter — name of the data center where the server is located; location_id — pool where the server is located; resource_id — server UUID; id — sensor ID; name — sensor name; type — sensor type
ipmi_fan_speed_rpm Gauge, RPM	Fan rotation speed	datacenter — name of the data center where the server is located; location_id — pool where the server is located; resource_id — server UUID; id — sensor ID; name — sensor name
ipmi_power_watts Gauge, watts	Power consumption	datacenter — name of the data center where the server is located; location_id — pool where the server is located; resource_id — server UUID; id — sensor ID; name — sensor name
ipmi_temperature_celsius Gauge, degrees °C	Server component temperature	datacenter — name of the data center where the server is located; location_id — pool where the server is located; resource_id — server UUID; name — sensor name; id — sensor ID
ipmi_voltage_volts Gauge, volts	Voltage	datacenter — name of the data center where the server is located; location_id — pool where the server is located; resource_id — server UUID; id — sensor ID; name — sensor name
ipmi_sensor_value Gauge	Data from the IPMI sensor. The unit of measurement depends on the sensor type	datacenter — name of the data center where the server is located; location_id — pool where the server is located; resource_id — server UUID; id — sensor ID; name — sensor name; type — sensor type

Cloud servers

The cloud servers namespace is compute.

Metrics are collected at the hypervisor level. Metrics from the hypervisor side and metrics inside the cloud server OS may differ in values:

• hypervisor metrics reflect the physical state of cloud servers and their resources — CPU, memory, disks, and network, including disk and network operation maintenance work and virtual machine service operations;
• in-OS metrics reflect only computations seen by the OS kernel and do not take background work into account.

Under high disk or network load, metric values may differ — this is expected behavior for a virtualized environment, not a metric collection error.

Metric collection does not depend on the OS state. Metrics are collected even if the cloud server is booting, is under heavy load, or is not responding over the network. Metrics are not collected only when the cloud server is powered off.

Metric name Type, unit of measurement	Description	Labels
server_meta_info Gauge, UNIT	Cloud server metadata	project_name — project name; server_uuid — cloud server ID; server_name — cloud server name
server_state_info Gauge, UNIT	Current cloud server status: 0 (no state) — server state is unknown or not defined; 1 (running) — server is running; 2 (blocked) — server is blocked (waiting for resources or completion of a disk read/write operation); 3 (paused) — server is paused; 4 (shutdown) — server is shutting down; 5 (shut off) — server is stopped and powered off; 6 (crashed) — server crashed; 7 (pmsuspended) — server is suspended and has entered sleep mode	server_uuid — cloud server ID
cpu_count Gauge, count	Number of cloud server vCPUs	server_uuid — cloud server ID
vcpu_usage_ratio Gauge, %	Utilization of each cloud server vCPU in %	server_uuid — cloud server ID; vcpu — virtual processor number
cpu_usage_ratio Gauge, %	Cloud server CPU utilization in %	server_uuid — cloud server ID
memory_total_bytes Gauge, bytes	Cloud server memory size	server_uuid — cloud server ID
memory_usage_bytes Gauge, bytes	Memory size used by the cloud server	server_uuid — cloud server ID
memory_usage_ratio Gauge, %	percentage of RAM used by the cloud server	server_uuid — cloud server ID
memory_major_faults_total Counter, count	Number of major page faults (major page faults) that have occurred on the cloud server since its last startup	server_uuid — cloud server ID
disk_meta_info Gauge, UNIT	Cloud server disk metadata	server_uuid — cloud server ID; disk_uuid — cloud server disk ID; disk_type="network" / "file" — disk type
disk_write_requests_rate Gauge, operations/s	Number of write operations per second on the cloud server disk	server_uuid — cloud server ID; disk_uuid — cloud server disk ID
disk_read_requests_rate Gauge, operations/s	Number of read operations per second from the cloud server disk	server_uuid — cloud server ID; disk_uuid — cloud server disk ID
disk_read_requests_limit Gauge, operations/s	Maximum number of read operations for the cloud server disk	server_uuid — cloud server ID; disk_uuid — cloud server disk ID
disk_write_requests_limit Gauge, operations/s	Maximum number of write operations for the cloud server disk	server_uuid — cloud server ID; disk_uuid — cloud server disk ID
disk_flush_requests_rate Gauge, ops/s	Number of disk cache flush operations for the cloud server disk	server_uuid — cloud server ID; disk_uuid — cloud server disk ID
disk_read_bytes_rate Gauge, bytes/s	Number of bytes per second read from the cloud server disk	server_uuid — cloud server ID; disk_uuid — cloud server disk ID
disk_write_bytes_rate Gauge, bytes/s	Number of bytes per second written to the cloud server disk	server_uuid — cloud server ID; disk_uuid — cloud server disk ID
network_meta_info Gauge, UNIT	Cloud server network interface metadata	server_uuid — cloud server ID; virtual_interface — cloud server port ID
network_receive_packets_rate Gauge, packets/s	Number of packets per second received on the cloud server network interface	server_uuid — cloud server ID; virtual_interface — cloud server port ID
network_transmit_packets_rate Gauge, packets/s	Number of packets per second transmitted through the cloud server network interface	server_uuid — cloud server ID; virtual_interface — cloud server port ID
network_receive_bits_rate Gauge, bits/s	Number of bits per second received on the cloud server network interface	server_uuid — cloud server ID; virtual_interface — cloud server port ID
network_transmit_bits_rate Gauge, bits/s	Number of bits per second sent through the cloud server network interface	server_uuid — cloud server ID; virtual_interface — cloud server port ID

Dedicated Inference

Dedicated Inference namespace — inference.

Metric name Type, unit of measurement	Description	Labels
info Gauge, UNIT	Inference service metadata. The metric value is always 1	service_id — inference service ID; name — inference service name; status — inference service status; availability_zone — pool segment where the inference service is located; region — pool where the inference service is located; gpu_count — number of GPUs; gpu_name — GPU model; ram_mb — RAM size in MB; vcpu — number of vCPUs; disk_size_gb — disk size in GB; disk_type — disk type; vm_type — configuration line of the cloud server on which the inference service is deployed: standard — Standard; hi-freq — HighFreq
config_info Gauge, UNIT	Configuration of the inference service. The metric value is always 1	service_id — inference service ID; max_concurrent_requests — maximum number of concurrent requests; public_access_enabled — connection type to the inference service: true — public access; false — private access; logs_enabled — logging: true — enabled; false — disabled
autoscaling_config_info Gauge, UNIT	Autoscaling settings of the inference service. The metric value is always 1	service_id — inference service ID; min_replicas — minimum number of inference instances; max_replicas — maximum number of inference instances
model_config_info Gauge, UNIT	Inference service model configuration. The metric value is always 1	service_id — inference service ID; model_id — model name; model_quantization — model quantization type; cache_quantization — data type for KV cache (Key-Value Cache); context_window_size — maximum context length
vllm_cache_config_info Gauge, UNIT	vLLM engine KV cache settings. The metric value is always 1	service_id — inference service ID; region — pool where the inference service is located; pod — inference instance; block_size — KV cache block size in tokens; cache_dtype — KV cache data type; cpu_kvcache_space_bytes — CPU space for KV cache, bytes; cpu_offload_gb — additional CPU memory size for KV cache in GB; enable_prefix_caching — request prefix caching: true — enabled; false — disabled; gpu_memory_utilization — allowable GPU memory utilization fraction; kv_cache_memory_bytes — GPU memory size for KV cache in bytes; kv_offloading_size — size of the area for temporarily offloading KV cache to RAM in bytes; kv_sharing_fast_prefill — accelerated processing of cached request prefixes: true — enabled; false — disabled; num_cpu_blocks — number of KV cache blocks in CPU memory; num_gpu_blocks — number of KV cache blocks in GPU memory; num_gpu_blocks_override — explicitly set number of KV cache blocks in GPU memory; prefix_caching_hash_algo — request prefix hashing algorithm; sliding_window — sliding context window size in tokens; swap_space — KV cache swap space size in bytes
vllm_kv_cache_usage_perc Gauge, fraction	Fraction of used KV cache blocks (from 0 to 1)	engine — unique identifier of the vLLM engine; model_name — model name; service_id — inference service ID; region — pool where the inference service is located; pod — inference instance
vllm_prefix_cache_hits_total Gauge, pcs	Number of prefix cache hits	engine — unique identifier of the vLLM engine; model_name — model name; service_id — inference service ID; region — pool where the inference service is located; pod — inference instance
vllm_prefix_cache_queries_total Gauge, pcs	Number of queries to the prefix cache	engine — unique identifier of the vLLM engine; model_name — model name; service_id — inference service ID; region — pool where the inference service is located; pod — inference instance
vllm_prompt_tokens_total Gauge, tokens	Total number of prompt tokens processed by the model	engine — unique identifier of the vLLM engine; model_name — model name; service_id — inference service ID; region — pool where the inference service is located; pod — inference instance
vllm_generation_tokens_total Gauge, tokens	Total number of generated tokens	engine — unique identifier of the vLLM engine; model_name — model name; service_id — inference service ID; region — pool where the inference service is located; pod — inference instance
vllm_num_requests_running Gauge, pcs	Number of requests currently being processed by the model	engine — unique identifier of the vLLM engine; model_name — model name; service_id — inference service ID; region — pool where the inference service is located; pod — inference instance
vllm_num_requests_waiting Gauge, pcs	Number of requests waiting in the queue	engine — unique identifier of the vLLM engine; model_name — model name; service_id — inference service ID; region — pool where the inference service is located; pod — inference instance
vllm_request_success_total Gauge, pcs	Number of successfully processed requests	engine — unique identifier of the vLLM engine; model_name — model name; service_id — inference service ID; region — pool where the inference service is located; pod — inference instance; finished_reason — reason for completing the request processing
vllm_engine_sleep_state Gauge, bool	Inference instance load state. The metric exports three lines at once — one line for each value of the sleep_state label. The value 1 is assigned to the line with the current state of the inference instance. The other lines are assigned the value 0: sleep_state="awake" — inference instance is active and ready to process requests; sleep_state="weights_offloaded" — model weights are offloaded from GPU memory, inference instance is partially freed; sleep_state="discard_all" — model weights and KV cache data are discarded, inference instance is completely freed	engine — unique identifier of the vLLM engine; model_name — model name; service_id — inference service ID; region — pool where the inference service is located; pod — inference instance; sleep_state — inference instance state: awake, weights_offloaded, or discard_all
vllm_e2e_request_latency_seconds Histogram, seconds	Request processing latency from receiving the prompt to full completion of the response (end-to-end)	le — upper bound of the histogram bucket; engine — unique identifier of the vLLM engine; model_name — model name; service_id — inference service ID; region — pool where the inference service is located; pod — inference instance
vllm_request_queue_time_seconds Histogram, seconds	Request wait time in the queue	le — upper bound of the histogram bucket; engine — unique identifier of the vLLM engine; model_name — model name; service_id — inference service ID; region — pool where the inference service is located; pod — inference instance
vllm_request_prefill_time_seconds Histogram, seconds	Processing time of the input prompt	le — upper bound of the histogram bucket; engine — unique identifier of the vLLM engine; model_name — model name; service_id — inference service ID; region — pool where the inference service is located; pod — inference instance
vllm_request_decode_time_seconds Histogram, seconds	Response generation time for a request	le — upper bound of the histogram bucket; engine — unique identifier of the vLLM engine; model_name — model name; service_id — inference service ID; region — pool where the inference service is located; pod — inference instance
vllm_request_inference_time_seconds Histogram, seconds	Total prompt processing and response generation time	le — upper bound of the histogram bucket; engine — unique identifier of the vLLM engine; model_name — model name; service_id — inference service ID; region — pool where the inference service is located; pod — inference instance
vllm_request_time_per_output_token_seconds Histogram, seconds	Average time to generate each token	le — upper bound of the histogram bucket; engine — unique identifier of the vLLM engine; model_name — model name; service_id — inference service ID; region — pool where the inference service is located; pod — inference instance
vllm_time_to_first_token_seconds Histogram, seconds	Time from receiving the request to generating the first token	le — upper bound of the histogram bucket; engine — unique identifier of the vLLM engine; model_name — model name; service_id — inference service ID; region — pool where the inference service is located; pod — inference instance
vllm_inter_token_latency_seconds Histogram, seconds	Inter-token latency	le — upper bound of the histogram bucket; engine — unique identifier of the vLLM engine; model_name — model name; service_id — inference service ID; region — pool where the inference service is located; pod — inference instance
vllm_request_prompt_tokens Histogram, tokens	Number of input tokens per request	le — upper bound of the histogram bucket; engine — unique identifier of the vLLM engine; model_name — model name; service_id — inference service ID; region — pool where the inference service is located; pod — inference instance
vllm_request_generation_tokens Histogram, tokens	Number of generated tokens per request	le — upper bound of the histogram bucket; engine — unique identifier of the vLLM engine; model_name — model name; service_id — inference service ID; region — pool where the inference service is located; pod — inference instance
vllm_iteration_tokens_total Histogram, tokens	Number of tokens processed by the model per iteration across all active requests	le — upper bound of the histogram bucket; engine — unique identifier of the vLLM engine; model_name — model name; service_id — inference service ID; region — pool where the inference service is located; pod — inference instance
cpu_usage_cores Gauge, cores	Current CPU consumption by the inference instance	service_id — inference service ID; region — pool where the inference service is located; pod — inference instance
cpu_limit_cores Gauge, cores	Configured CPU limit for the inference instance	service_id — inference service ID; region — pool where the inference service is located; pod — inference instance
memory_usage_bytes Gauge, bytes	Current memory consumption by the inference instance	service_id — inference service ID; region — pool where the inference service is located; pod — inference instance
memory_limit_bytes Gauge, bytes	Configured memory limit for the inference instance	service_id — inference service ID; region — pool where the inference service is located; pod — inference instance
http_requests_total Gauge, pcs	Total number of HTTP requests to the inference service	service_id — inference service ID; region — pool where the inference service is located; code — HTTP request response code
http_requests_bytes_total Gauge, bytes	Total size of HTTP requests	service_id — inference service ID; region — pool where the inference service is located
http_responses_bytes_total Gauge, bytes	Total size of HTTP responses	service_id — inference service ID; region — pool where the inference service is located
network_receive_bytes_total Gauge, bytes/s	Number of bytes per second received by the inference instance	service_id — inference service ID; region — pool where the inference service is located; pod — inference instance
network_receive_packets_total Gauge, packets/s	Number of packets per second received by the inference instance	service_id — inference service ID; region — pool where the inference service is located; pod — inference instance
network_receive_packets_dropped_total Gauge, packets/s	Number of packets per second dropped on receive by the inference instance	service_id — inference service ID; region — pool where the inference service is located; pod — inference instance
network_transmit_bytes_total Gauge, bytes/s	Number of bytes per second transmitted by the inference instance	service_id — inference service ID; region — pool where the inference service is located; pod — inference instance
network_transmit_packets_total Gauge, packets/s	Number of packets per second transmitted by the inference instance	service_id — inference service ID; region — pool where the inference service is located; pod — inference instance
network_transmit_packets_dropped_total Gauge, packets/s	Number of packets per second dropped on transmit by the inference instance	service_id — inference service ID; region — pool where the inference service is located; pod — inference instance