Metrics Semantic Conventions
The following semantic conventions surrounding metrics are defined:
- General Guidelines: General metrics guidelines.
- Database: For SQL and NoSQL client metrics.
- FaaS: For Function as a Service metrics.
- HTTP: For HTTP client and server metrics.
- RPC: For RPC client and server metrics.
- System metrics
Metric names and attributes exist within a single universe and a single hierarchy. Metric names and attributes MUST be considered within the universe of all existing metric names. When defining new metric names and attributes, consider the prior art of existing standard metrics and metrics from frameworks/libraries.
Associated metrics SHOULD be nested together in a hierarchy based on their usage. Define a top-level hierarchy for common metric categories: for OS metrics, like CPU and network; for app runtimes, like GC internals. Libraries and frameworks should nest their metrics into a hierarchy as well. This aids in discovery and adhoc comparison. This allows a user to find similar metrics given a certain metric.
The hierarchical structure of metrics defines the namespacing. Supporting OpenTelemetry artifacts define the metric structures and hierarchies for some categories of metrics, and these can assist decisions when creating future metrics.
Common attributes SHOULD be consistently named. This aids in discoverability and disambiguates similar attributes to metric names.
“As a rule of thumb, aggregations over all the attributes of a given metric SHOULD be meaningful,” as Prometheus recommends.
Semantic ambiguity SHOULD be avoided. Use prefixed metric names in cases
where similar metrics have significantly different implementations across the
breadth of all existing metrics. For example, every garbage collected runtime
has slightly different strategies and measures. Using a single set of metric
names for GC, not divided by the runtime, could create dissimilar comparisons
and confusion for end users. (For example, prefer
process.runtime.gc.*.) Measures of many operating system metrics are similarly
A new metric MUST NOT be added with the same name as a metric that existed in the past but was renamed (with a corresponding schema file).
When introducing a new metric name check all existing schema files to make sure the name does not appear as a key of any “rename_metrics” section (keys denote old metric names in rename operations).
Conventional metrics or metrics that have their units included in
OpenTelemetry metadata (e.g.
metric.WithUnit in Go) SHOULD NOT include the
units in the metric name. Units may be included when it provides additional
meaning to the metric name. Metrics MUST, above all, be understandable and
When building components that interoperate between OpenTelemetry and a system using the OpenMetrics exposition format, use the OpenMetrics Guidelines.
Metric namespaces SHOULD NOT be pluralized.
Metric names SHOULD NOT be pluralized, unless the value being recorded
represents discrete instances of a
Generally, the name SHOULD be pluralized only if the unit of the metric in
question is a non-unit (like
system.cpu.timeshould not be pluralized, even if many data points are recorded.
system.network.packetsshould be pluralized, even if only a single data point is recorded.
If the value being recorded represents the count of concepts signified
by the namespace then the metric should be named
count (within its namespace).
For example if we have a namespace
system.process which contains all metrics related
to the processes then to represent the count of the processes we can have a metric named
UpDownCounters SHOULD NOT use
_total because then they will look like
Counters SHOULD NOT append
_total either because then their meaning will
be confusing in delta backends.
The following semantic conventions aim to keep naming consistent. They provide guidelines for most of the cases in this specification and should be followed for other instruments not explicitly defined in this document.
limit - an instrument that measures the constant, known total amount of something should be called
entity.limit. For example,
system.memory.limitfor the total amount of memory on a system.
usage - an instrument that measures an amount used out of a known total (limit) amount should be called
entity.usage. For example,
state = used | cached | free | ...for the amount of memory in a each state. Where appropriate, the sum of usage over all attribute values SHOULD be equal to the limit.
A measure of the amount consumed of an unlimited resource, or of a resource whose limit is unknowable, is differentiated from usage. For example, the maximum possible amount of virtual memory that a process may consume may fluctuate over time and is not typically known.
utilization - an instrument that measures the fraction of usage out of its limit should be called
entity.utilization. For example,
system.memory.utilizationfor the fraction of memory in use. Utilization can be with respect to a fixed limit or a soft limit. Utilization values are represended as a ratio and are typically in the range
[0, 1], but may go above 1 in case of exceeding a soft limit.
time - an instrument that measures passage of time should be called
entity.time. For example,
state = idle | user | system | .... time measurements are not necessarily wall time and can be less than or greater than the real wall time between measurements.
time instruments are a special case of usage metrics, where the limit can usually be calculated as the sum of time over all attribute values. utilization for time instruments can be derived automatically using metric event timestamps. For example,
system.cpu.utilizationis defined as the difference in
system.cpu.timemeasurements divided by the elapsed time and number of CPUs.
io - an instrument that measures bidirectional data flow should be called
entity.ioand have attributes for direction. For example,
Other instruments that do not fit the above descriptions may be named more freely. For example,
system.network.packets. Units do not need to be specified in the names since they are included during instrument creation, but can be added if there is ambiguity.
Units should follow the Unified Code for Units of Measure.
- Instruments for utilization metrics (that measure the fraction out of a
total) are dimensionless and SHOULD use the default unit
- All non-units that use curly braces to annotate a quantity need to match the
grammatical number of the quantity it represent. For example if measuring the
number of individual requests to a process the unit would be
- Instruments that measure an integer count of something SHOULD only use
annotations with curly braces to
give additional meaning without the leading default unit (
1). For example, use
- Instrument units other than
1and those that use annotations SHOULD be specified using the UCUM case sensitive (“c/s”) variant. For example, “Cel” for the unit with full name “degree Celsius”.
- Instruments SHOULD use non-prefixed units (i.e.
MiBy) unless there is good technical reason to not do so.
- When instruments are measuring durations, seconds (i.e.
s) SHOULD be used.
The semantic metric conventions specification is written to use the names of the synchronous instrument types,
UpDownCounter. However, compliant implementations MAY use the asynchronous equivalent instead,
Asynchronous Counter or
Whether implementations choose the synchronous type or the asynchronous equivalent is considered to be an
implementation detail. Both choices are compliant with this specification.
UpDownCounter metrics, the same attribute values used to record an increment SHOULD be used to record
any associated decrement, otherwise those increments and decrements will end up as different timeseries.
For example, if you are tracking
active_requests with an
UpDownCounter, and you are incrementing it each time a
request starts and decrementing it each time a request ends, then any attributes which are not yet available when
incrementing the counter at request start should not be used when decrementing the counter at request end.