> ## Documentation Index
> Fetch the complete documentation index at: https://otel.fyi/llms.txt
> Use this file to discover all available pages before exploring further.

# Transform

> OpenTelemetry processor for Transform

# Transform Processor

![Status](https://img.shields.io/badge/status-beta-yellow)

**Available in:** `contrib`, `k8s`

**Maintainers:** [@TylerHelmuth](https://github.com/TylerHelmuth), [@evan-bradley](https://github.com/evan-bradley), [@edmocosta](https://github.com/edmocosta), [@bogdandrutu](https://github.com/bogdandrutu)

**Source:** [opentelemetry-collector-contrib](https://github.com/open-telemetry/opentelemetry-collector-contrib/tree/main/processor/transformprocessor)

## Supported Telemetry

![Logs](https://img.shields.io/badge/logs-beta-blue) ![Metrics](https://img.shields.io/badge/metrics-beta-green) ![Traces](https://img.shields.io/badge/traces-beta-orange)

## Overview

The Transform Processor modifies telemetry based on configuration using the [OpenTelemetry Transformation Language](https://github.com/open-telemetry/opentelemetry-collector-contrib/tree/main/pkg/ottl) (OTTL).

For each signal type, the processor takes a list of statements and executes them against the incoming telemetry, following the order specified in the configuration.
Each statement can access and transform telemetry using functions, and allows the use of a condition to help decide whether the function should be executed.

* [Config](#config)
* [Grammar](#grammar)
* [Supported functions](#supported-functions)
* [Examples](#examples)
* [Troubleshooting](#troubleshooting)
* [Contributing](#contributing)
* [Feature Gate](#feature-gate)
* [Compatibility](#compatibility)
* [Performance and scaling](#performance-and-scaling)
* [Self-observability](#self-observability)

## Config

### General Config

> \[!NOTE]
> If you don't know how to write OTTL statements yet, first see OTTL's [Getting Started](https://github.com/open-telemetry/opentelemetry-collector-contrib/blob/main/pkg/ottl/README.md#getting-started) docs.

```yaml theme={null}
transform:
  error_mode: ignore
  <trace|metric|log|profile>_statements: []
```

The Transform Processor's primary configuration section is broken down by signal (traces, metrics, logs, and profiles)
and allows you to configure a list of statements for the processor to execute. The list can be made of:

* OTTL statements. This option will meet most user's needs. See [Basic Config](#basic-config) for more details.
* Objects, which allows users to apply configuration options to a specific list of statements. See [Advanced Config](#advanced-config) for more details.

> \[!NOTE]
> Support for the `profile` signal (the `profile_statements` configuration and the `profile` context) is experimental and subject to change or removal in the future.

Within each `<signal_statements>` list, only certain OTTL Path prefixes can be used:

| Signal | Path Prefix Values |
| - | - |
| trace\_statements | `resource`, `scope`, `span`, and `spanevent` |
| metric\_statements | `resource`, `scope`, `metric`, `datapoint`, and `exemplar` |
| log\_statements | `resource`, `scope`, and `log` |
| profile\_statements | `resource`, `scope`, and `profile` |

This means, for example, that you cannot use the Path `span.attributes` within the `log_statements` configuration section.

`error_mode`: determines how the processor treats errors that occur while processing a statement.
If the top-level `error_mode` is not specified, `ignore` will be used.
The top-level `error_mode` can be overridden at statement group level, offering more granular control over error handling. If the statement group `error_mode` is not specified, the top-level `error_mode` is applied.

| error\_mode | description |
| - | - |
| ignore | The processor ignores errors returned by statements, logs the error, and continues on to the next statement.  This is the recommended mode. |
| silent | The processor ignores errors returned by statements, does not log the error, and continues on to the next statement. |
| propagate | The processor returns the error up the pipeline.  This will result in the payload being dropped from the collector. |

### Basic Config

> \[!NOTE]
> If you don't know how to write OTTL statements yet, first see OTTL's [Getting Started](https://github.com/open-telemetry/opentelemetry-collector-contrib/blob/main/pkg/ottl/README.md#getting-started) docs.

The basic configuration style allows you to configure OTTL statements as a list,
without worrying about extra configurations.

**This is the simplest way to configure the Transform Processor.** If you need global conditions or specific error modes see [Advanced Config](#advanced-config).

Format:

```yaml theme={null}
transform:
  error_mode: ignore
  <trace|metric|log|profile>_statements:
    - string
    - string
    - string
```

Example:

```yaml theme={null}
transform:
  error_mode: ignore
  trace_statements:
    - keep_keys(span.attributes, ["service.name", "service.namespace", "cloud.region", "process.command_line"])
    - replace_pattern(span.attributes["process.command_line"], "password\\=[^\\s]*(\\s?)", "password=***")
    - limit(span.attributes, 100, [])
    - truncate_all(span.attributes, 4096)
  metric_statements:
    - keep_keys(resource.attributes, ["host.name"])
    - truncate_all(resource.attributes, 4096)
    - set(metric.description, "Sum") where metric.type == "Sum"
    - convert_sum_to_gauge() where metric.name == "system.processes.count"
    - convert_gauge_to_sum("cumulative", false) where metric.name == "prometheus_metric"
  log_statements:
    - set(log.severity_text, "FAIL") where log.body == "request failed"
    - replace_all_matches(log.attributes, "/user/*/list/*", "/user/{userId}/list/{listId}")
    - replace_all_patterns(log.attributes, "value", "/account/\\d{4}", "/account/{accountId}")
    - set(log.body, log.attributes["http.route"])
  profile_statements:
    - keep_keys(resource.attributes, ["host.name"])
    - set(profile.attributes["tag"], "profile#23")
    - set(profile.original_payload_format, "json")
```

In some situations a combination of Paths, functions, or enums is not allowed, and the solution
might require multiple [Advanced Config](#advanced-config) configuration groups.
See [Context Inference](#context-inference) for more details.

### Advanced Config

> \[!NOTE]
> If you don't know how to write OTTL statements yet, first see OTTL's [Getting Started](https://github.com/open-telemetry/opentelemetry-collector-contrib/blob/main/pkg/ottl/README.md#getting-started) docs.

For more complex use cases you may need to use the Transform Processor's
advanced configuration style to group related OTTL statements.

Format:

```yaml theme={null}
transform:
  error_mode: ignore
  <trace|metric|log|profile>_statements:
    - context: string
      error_mode: propagate
      shared_cache: true
      conditions: 
        - string
        - string
      statements:
        - string
        - string
        - string
    - context: string
      error_mode: silent
      statements:
        - string
        - string
        - string
```

`error_mode`: allows overriding the top-level `error_mode`. See [General Config](#general-config) for details on how to configure `error_mode`.

`shared_cache` (experimental): enables sharing per-context caches between all advanced statement groups with the option enabled in this Transform Processor instance. Use this when you want to hold state across multiple iterations over a set of data.

`conditions`: a list comprised of multiple where clauses, which will be processed as global conditions for the accompanying set of statements. The conditions are ORed together, which means only one condition needs to evaluate to true in order for the statements (including their individual Where clauses) to be executed.

`statements`: a list of OTTL statements.

Example:

```yaml theme={null}
transform:
  error_mode: ignore
  metric_statements:
    - error_mode: propagate
      conditions:
        - metric.type == METRIC_DATA_TYPE_SUM
      statements:
        - set(metric.description, "Sum")

  log_statements:
    - conditions:
        - IsMap(log.body) and log.body["object"] != nil
      statements:
        - set(log.body, log.attributes["http.route"])
```

The Transform Processor will enforce that all the Paths, functions, and enums used in a group's `statements` are parsable.
In some situations a combination of Paths, functions, or enums is not allowed, and it might require multiple configuration groups.
See [Context Inference](#context-inference) for more details.

### Context inference

> \[!NOTE]
> This is an advanced topic and is not necessary to get started using the Transform Processor.
> Read on if you're interested in how the Transform Processor parses your OTTL statements.

An OTTL Context defines which Paths, functions, and enums are available when parsing the statement.
The Transform Processor automatically infers the OTTL Context based on the paths defined in a statement.

This inference is based on the Path names, functions, and enums present in the statements.

The inference happens automatically because Path names are prefixed with the Context name. For example:

```yaml theme={null}
metric_statements:
  - set(metric.description, "test passed") where datapoint.attributes["test"] == "pass"
```

In this configuration, the inferred Context value is `datapoint`, as it is the only Context
that supports parsing both `datapoint` and `metric` Paths.

In the following example, the inferred Context is `metric`,
as `metric` is the context capable of parsing both `metric` and `resource` data.

```yaml theme={null}
metric_statements:
  - set(resource.attributes["test"], "passed")
  - set(metric.description, "test passed")
```

The primary benefit of context inference is that it enhances the efficiency of statement processing
by linking them to the most suitable context. This optimization ensures that data transformations
are both accurate and performant, leveraging the hierarchical structure of contexts to avoid unnecessary
iterations and improve overall processing efficiency.
All of this happens automatically, leaving you to write OTTL statements without worrying about Context.

In some situations a combination of Paths, functions, or enums is not allowed. For example:

```yaml theme={null}
metric_statements:
  - convert_sum_to_gauge() where metric.name == "system.processes.count"
  - limit(datapoint.attributes, 100, ["host.name"])
```

In this configuration, the `datapoint` Path prefixed is used in the same group of statements as the `convert_sum_to_gauge`
function. Since `convert_sum_to_gauge` can only be used with the metrics, not datapoints, but the list
statements contains a reference to the datapoints via the `datapoint` Path prefix, the group of statements cannot
be parsed.

The solution is to separate the statements into separate [Advanced Config](#advanced-config) groups:

```yaml theme={null}
metric_statements:
  - statements:
    - convert_sum_to_gauge() where metric.name == "system.processes.count"
  - statements:
      - limit(datapoint.attributes, 100, ["host.name"])
```

## Grammar

You can learn more in-depth details on the capabilities and limitations of the OpenTelemetry Transformation Language used by the Transform Processor by reading about its [grammar](https://github.com/open-telemetry/opentelemetry-collector-contrib/blob/main/pkg/ottl/LANGUAGE.md).

## Supported functions:

These common functions can be used for any Signal.

* [OTTL Functions](https://github.com/open-telemetry/opentelemetry-collector-contrib/tree/main/pkg/ottl/ottlfuncs)

In addition to the common OTTL functions, the processor defines its own functions to help with transformations specific to this processor:

**Metrics only functions**

* [convert\_sum\_to\_gauge](#convert_sum_to_gauge)
* [convert\_gauge\_to\_sum](#convert_gauge_to_sum)
* [extract\_avg\_metric](#extract_avg_metric)
* [extract\_count\_metric](#extract_count_metric)
* [extract\_percentile\_metric](#extract_percentile_metric)
* [extract\_sum\_metric](#extract_sum_metric)
* [convert\_summary\_count\_val\_to\_sum](#convert_summary_count_val_to_sum)
* [convert\_summary\_quantile\_val\_to\_gauge](#convert_summary_quantile_val_to_gauge)
* [convert\_summary\_sum\_val\_to\_sum](#convert_summary_sum_val_to_sum)
* [copy\_metric](#copy_metric)
* [scale\_metric](#scale_metric)
* [aggregate\_on\_attributes](#aggregate_on_attributes)
* [convert\_exponential\_histogram\_to\_histogram](#convert_exponential_histogram_to_histogram)
* [aggregate\_on\_attribute\_value](#aggregate_on_attribute_value)
* [merge\_histogram\_buckets](#merge_histogram_buckets)

**Logs only functions**

* [ParseCEF](#parsecef)
* [ParseCLF](#parseclf)
* [ParseELF](#parseelf)
* [ParseLEEF](#parseleef)

**Traces only functions**

* [set\_semconv\_span\_name](#set_semconv_span_name)

### convert\_sum\_to\_gauge

`convert_sum_to_gauge()`

Converts incoming metrics of type "Sum" to type "Gauge", retaining the metric's datapoints. Noop for metrics that are not of type "Sum".

**NOTE:** This function may cause a metric to break semantics for [Gauge metrics](https://github.com/open-telemetry/opentelemetry-specification/blob/main/specification/metrics/data-model.md#gauge). Use at your own risk.

Examples:

* `convert_sum_to_gauge()`

### convert\_gauge\_to\_sum

`convert_gauge_to_sum(aggregation_temporality, is_monotonic)`

Converts incoming metrics of type "Gauge" to type "Sum", retaining the metric's datapoints and setting its aggregation temporality and monotonicity accordingly. Noop for metrics that are not of type "Gauge".

`aggregation_temporality` is a string (`"cumulative"` or `"delta"`) that specifies the resultant metric's aggregation temporality. `is_monotonic` is a boolean that specifies the resultant metric's monotonicity.

**NOTE:** This function may cause a metric to break semantics for [Sum metrics](https://github.com/open-telemetry/opentelemetry-specification/blob/main/specification/metrics/data-model.md#sums). Use at your own risk.

Examples:

* `convert_gauge_to_sum("cumulative", false)`

* `convert_gauge_to_sum("delta", true)`

### extract\_avg\_metric

> \[!NOTE]\
> This function supports Histograms, ExponentialHistograms and Summaries.

`extract_avg_metric(Optional[suffix])`

The `extract_avg_metric` function creates a new Gauge metric from a Histogram, ExponentialHistogram or Summary by dividing the sum value by the count value for each data point. Data points where the sum value is missing or the count value is `0` are skipped, to avoid division by zero. A metric will only be created if there is at least one data point.

`suffix` is an optional string that defines the suffix for the metric name. By default, it is set to `_avg`.

The name for the new metric will be `<original metric name><suffix>`. The fields that are copied are: `timestamp`, `starttimestamp`, `attributes`, `description`, and `unit`.

The new metric that is created will be passed to all subsequent statements in the metrics statements list.

Examples:

* `extract_avg_metric()`

* `extract_avg_metric(".avg")`

### extract\_count\_metric

> \[!NOTE]\
> This function supports Histograms, ExponentialHistograms and Summaries.

`extract_count_metric(is_monotonic, Optional[suffix])`

The `extract_count_metric` function creates a new Sum metric from a Histogram, ExponentialHistogram or Summary's count value. A metric will only be created if there is at least one data point.

`is_monotonic` is a boolean representing the monotonicity of the new metric.
`suffix` is an optional string that defines the suffix for the metric name. By default, it is set to `_count`.

The name for the new metric will be `<original metric name><suffix>`. The fields that are copied are: `timestamp`, `starttimestamp`, `attributes`, `description`, and `aggregation_temporality`. As metrics of type Summary don't have an `aggregation_temporality` field, this field will be set to `AGGREGATION_TEMPORALITY_CUMULATIVE` for those metrics.

The new metric that is created will be passed to all subsequent statements in the metrics statements list.

> \[!WARNING]\
> This function may cause a metric to break semantics for [Sum metrics](https://github.com/open-telemetry/opentelemetry-specification/blob/main/specification/metrics/data-model.md#sums). Use only if you're confident you know what the resulting monotonicity should be.

Examples:

* `extract_count_metric(true, ".count")`

* `extract_count_metric(false, ".count")`

### extract\_percentile\_metric

> \[!NOTE]\
> This function supports Histograms and ExponentialHistograms.

`extract_percentile_metric(percentile, Optional[suffix])`

The `extract_percentile_metric` function creates a new Gauge metric from a Histogram or ExponentialHistogram by calculating the specified percentile value from the bucket counts. A metric will only be created if there is at least one data point.

`percentile` is a float64 value greater than 0 and less than 100 representing the desired percentile to extract (e.g., 50 for median, 95 for p95, 99 for p99).

`suffix` is an optional string that defines the suffix for the metric name. By default, it is set to `_p{percentile}` (e.g., `_p50`, `_p95`, `_p99`).

The name for the new metric will be `<original metric name><suffix>`. The fields that are copied are: `timestamp`, `starttimestamp`, `attributes`, `description`, and `unit`.

For Histograms, the function uses linear interpolation within buckets to estimate the percentile value. Since the lowest bucket has no finite lower bound (`-Inf`), the function uses the data point's `Min` value when available. If `Min` is not set and the upper bound of that bucket is non-negative, the lower bound is assumed to be `0`. Similarly, if the percentile falls in the last bucket (`+Inf` upper bound), the data point's `Max` value is used for interpolation when available. For ExponentialHistograms, it uses logarithmic interpolation appropriate for the exponential bucket structure.

The new metric that is created will be passed to all subsequent statements in the metrics statements list.

Examples:

* `extract_percentile_metric(50.0)` - Extract median (p50) with default suffix `_p50`

* `extract_percentile_metric(95.0, "_p95_custom")` - Extract p95 with custom suffix `_p95_custom`

* `extract_percentile_metric(99.9) where metric.name == "http.server.duration"` - Extract p99.9 only for specific metrics

### extract\_sum\_metric

> \[!NOTE]\
> This function supports Histograms, ExponentialHistograms and Summaries.

`extract_sum_metric(is_monotonic, Optional[suffix])`

The `extract_sum_metric` function creates a new Sum metric from a Histogram, ExponentialHistogram or Summary's sum value. If the sum value of a Histogram or ExponentialHistogram data point is missing, no data point is added to the output metric. A metric will only be created if there is at least one data point.

`is_monotonic` is a boolean representing the monotonicity of the new metric.
`suffix` is an optional string that defines the suffix for the metric name. By default, it is set to `_sum`.

The name for the new metric will be `<original metric name><suffix>`. The fields that are copied are: `timestamp`, `starttimestamp`, `attributes`, `description`, and `aggregation_temporality`. As metrics of type Summary don't have an `aggregation_temporality` field, this field will be set to `AGGREGATION_TEMPORALITY_CUMULATIVE` for those metrics.

The new metric that is created will be passed to all subsequent statements in the metrics statements list.

> \[!WARNING]\
> This function may cause a metric to break semantics for [Sum metrics](https://github.com/open-telemetry/opentelemetry-specification/blob/main/specification/metrics/data-model.md#sums). Use only if you're confident you know what the resulting monotonicity should be.

Examples:

* `extract_sum_metric(true, ".sum")`

* `extract_sum_metric(false, ".sum")`

### convert\_summary\_count\_val\_to\_sum

`convert_summary_count_val_to_sum(aggregation_temporality, is_monotonic, Optional[suffix])`

The `convert_summary_count_val_to_sum` function creates a new Sum metric from a Summary's count value.

`aggregation_temporality` is a string (`"cumulative"` or `"delta"`) representing the desired aggregation temporality of the new metric. `is_monotonic` is a boolean representing the monotonicity of the new metric.

`suffix` is an optional string that defines the suffix for the metric name. By default, it is set to `_count`.

The name for the new metric will be `<summary metric name><suffix>`. The fields that are copied are: `timestamp`, `starttimestamp`, `attributes`, and `description`. The new metric that is created will be passed to all functions in the metrics statements list.  Function conditions will apply.

**NOTE:** This function may cause a metric to break semantics for [Sum metrics](https://github.com/open-telemetry/opentelemetry-specification/blob/main/specification/metrics/data-model.md#sums). Use at your own risk.

Examples:

* `convert_summary_count_val_to_sum("delta", true, ".count")`

* `convert_summary_count_val_to_sum("cumulative", false, ".count")`

### convert\_summary\_quantile\_val\_to\_gauge

`convert_summary_quantile_val_to_gauge(Optional[attributeKey], Optional[suffix])`

The `convert_summary_quantile_val_to_gauge` function creates a new Gauge metric and injects each of the Summary's quantiles into a single Gauge datapoint.

`attributeKey` is an optional string that specifies the attribute key holding the quantile value for each corresponding output data point. The default key is `quantile`.
`suffix` is an optional string representing the suffix of the metric name. The default value is `.quantiles`.

The name for the new metric will be `<summary metric name>.quantiles`. The fields that are copied are: `timestamp`, `starttimestamp`, `attributes`, `unit` and `description`. The new metric that is created will be passed to all functions in the metrics statements list. Function conditions will apply.

Examples:

* `convert_summary_quantile_val_to_gauge("custom_quantile", "custom_suffix")`

* `convert_summary_quantile_val_to_gauge("custom_quantile")`

* `convert_summary_quantile_val_to_gauge()`

### convert\_summary\_sum\_val\_to\_sum

`convert_summary_sum_val_to_sum(aggregation_temporality, is_monotonic, Optional[suffix])`

The `convert_summary_sum_val_to_sum` function creates a new Sum metric from a Summary's sum value.

`aggregation_temporality` is a string (`"cumulative"` or `"delta"`) representing the desired aggregation temporality of the new metric. `is_monotonic` is a boolean representing the monotonicity of the new metric.
`suffix` is an optional string that defines the suffix for the metric name. By default, it is set to `_sum`.

The name for the new metric will be `<summary metric name><suffix>`. The fields that are copied are: `timestamp`, `starttimestamp`, `attributes`, and `description`. The new metric that is created will be passed to all functions in the metrics statements list.  Function conditions will apply.

**NOTE:** This function may cause a metric to break semantics for [Sum metrics](https://github.com/open-telemetry/opentelemetry-specification/blob/main/specification/metrics/data-model.md#sums). Use at your own risk.

Examples:

* `convert_summary_sum_val_to_sum("delta", true, ".sum")`

* `convert_summary_sum_val_to_sum("cumulative", false, ".sum")`

### copy\_metric

`copy_metric(Optional[name], Optional[description], Optional[unit])`

The `copy_metric` function copies the current metric, adding it to the end of the metric slice.

`name` is an optional string. `description` is an optional string. `unit` is an optional string.

The new metric will be exactly the same as the current metric.  You can use the optional parameters to set the new metric's name, description, and unit.

**NOTE:** The new metric is appended to the end of the metric slice and therefore will be included in all the metric statements. It is a best practice to ALWAYS include a Where clause when copying a metric that WILL NOT match the new metric.

Examples:

* `copy_metric(name="http.request.status_code", unit="s") where metric.name == "http.status_code`

* `copy_metric(desc="new desc") where metric.description == "old desc"`

### convert\_exponential\_histogram\_to\_histogram

**Warning:** This conversion is not defined by the OpenTelemetry Specification and may lose precision.

`convert_exponential_histogram_to_histogram(distribution, [ExplicitBounds])`

Converts an exponential histogram to an explicit histogram.

The function requires two arguments:

* `distribution` controls how each exponential bucket is allocated:
  * `upper` places the bucket count at its upper bound.
  * `midpoint` places the bucket count at its arithmetic midpoint.
  * `uniform` assumes a uniform distribution within the bucket and allocates counts according to the overlap with each explicit bucket. Rounding is deterministic.
  * `random` uses the same overlap calculation as `uniform`, with randomized rounding.
* `ExplicitBounds` is a non-empty, strictly increasing list of explicit histogram bounds.

The result:

* Converts positive, negative, and zero buckets.
* Preserves the total count.
* Contains `len(ExplicitBounds) + 1` bucket counts. For bounds `[b0, b1]`, the buckets are `(-Inf, b0]`, `(b0, b1]`, and `(b1, +Inf)`.

For example:

```ottl theme={null}
convert_exponential_histogram_to_histogram("random", [0.0, 10.0, 100.0, 1000.0, 10000.0])
```

Use this function only when downstream consumers cannot accept exponential histograms.

### scale\_metric

`scale_metric(factor, Optional[unit])`

The `scale_metric` function multiplies the values in the data points in the metric by the float value `factor`.
If the optional string `unit` is provided, the metric's unit will be set to this value.
The supported data types are:

Supported metric types are `Gauge`, `Sum`, `Histogram`, and `Summary`.

Examples:

* `scale_metric(0.1)`: Scale the metric by a factor of `0.1`. The unit of the metric will not be modified.
* `scale_metric(10.0, "kWh")`: Scale the metric by a factor of `10.0` and sets the unit to `kWh`.

### aggregate\_on\_attributes

`aggregate_on_attributes(function, Optional[attributes])`

The `aggregate_on_attributes` function aggregates all data points in the metric.

`function` is a case-sensitive string that represents the aggregation function to apply. `attributes` is an optional list of datapoint attribute keys (strings) to aggregate upon.

The optional `attributes` argument controls which datapoint attributes are kept before aggregation:

* omitted (`aggregate_on_attributes(function)`): all existing data point attributes are kept. Aggregation only combines data points that already share the same full attribute set.
* `[]` (empty list): all datapoint attributes are removed. Aggregation is performed across all data points in the metric.
* `[...]` (list of attribute keys): only the specified data point attributes are kept. All other attributes are removed, and aggregation is performed by grouping on the remaining attributes.

**NOTE:** This function is supported only in `metric` context.

The following metric types can be aggregated:

* sum
* gauge
* histogram
* exponential histogram

Supported aggregation functions are:

* sum
* max
* min
* mean
* median
* count

**NOTE:** Only the `sum` aggregation function is supported for histogram and exponential histogram datatypes.

Examples:

* `aggregate_on_attributes("sum", ["attr1", "attr2"]) where metric.name == "system.memory.usage"`
* `aggregate_on_attributes("max") where metric.name == "system.memory.usage"`
* `aggregate_on_attributes("max", []) where metric.name == "system.memory.usage"`

The `aggregate_on_attributes` function can also be used in conjunction with
[keep\_matching\_keys](https://github.com/open-telemetry/opentelemetry-collector-contrib/blob/main/pkg/ottl/ottlfuncs#keep_matching_keys) or
[delete\_matching\_keys](https://github.com/open-telemetry/opentelemetry-collector-contrib/blob/main/pkg/ottl/ottlfuncs#delete_matching_keys).

For example, to remove attribute keys matching a regex and aggregate the metrics on the remaining attributes, you can perform the following statement sequence:

```yaml theme={null}
statements:
   - delete_matching_keys(resource.attributes, "(?i).*myRegex.*") where metric.name == "system.memory.usage"
   - aggregate_on_attributes("sum") where metric.name == "system.memory.usage"
```

To aggregate only using a specified set of attributes, you can use `keep_matching_keys`.

### aggregate\_on\_attribute\_value

`aggregate_on_attribute_value(function, attribute, values, newValue)`

The `aggregate_on_attribute_value` function aggregates all datapoints in the metric containing the attribute `attribute` (type string) with one of the values present in the `values` parameter (list of strings) into a single datapoint where the attribute has the value `newValue` (type string). `function` is a case-sensitive string that represents the aggregation function.

**NOTE:** This function is supported only in `metric` context.

The following metric types can be aggregated:

* sum
* gauge
* histogram
* exponential histogram

Supported aggregation functions are:

* sum
* max
* min
* mean
* median
* count

**NOTE:** Only the `sum` aggregation function is supported for histogram and exponential histogram datatypes.

Examples:

* `aggregate_on_attribute_value("sum", "attr1", ["val1", "val2"], "new_val") where metric.name == "system.memory.usage"`

The `aggregate_on_attribute_value` function can also be used in conjunction with
[keep\_matching\_keys](https://github.com/open-telemetry/opentelemetry-collector-contrib/blob/main/pkg/ottl/ottlfuncs#keep_matching_keys) or
[delete\_matching\_keys](https://github.com/open-telemetry/opentelemetry-collector-contrib/blob/main/pkg/ottl/ottlfuncs#delete_matching_keys).

For example, to remove attribute keys matching a regex and aggregate the metrics on the remaining attributes, you can perform the following statement sequence:

```yaml theme={null}
statements:
   - delete_matching_keys(resource.attributes, "(?i).*myRegex.*") where metric.name == "system.memory.usage"
   - aggregate_on_attribute_value("sum", "attr1", ["val1", "val2"], "new_val") where metric.name == "system.memory.usage"
```

To aggregate only using a specified set of attributes, you can use `keep_matching_keys`.

### merge\_histogram\_buckets

`merge_histogram_buckets(target_value, method)`

The `merge_histogram_buckets` function merges explicit histogram buckets. The `method` argument is optional and defaults to `remove_explicit_bound`.

`target_value` is interpreted according to `method`:

* `remove_explicit_bound`: `target_value` is the explicit boundary to remove. The function merges the bucket ending at this boundary with the next bucket. This method uses floating-point tolerance (epsilon = 1e-12) when matching the boundary.
* `limit_buckets`: `target_value` is the maximum number of buckets to keep. It must be a positive integer. The function reduces resolution with a single uniform compaction pass. It chooses the smallest divisor that keeps the resulting bucket count at or below `target_value`, merges adjacent buckets in groups of that size from lower to higher bucket order, combines their counts, and keeps any partial final group. Bucket count values and boundary widths do not affect which buckets are merged. The resulting histogram may have fewer than `target_value` buckets when no smaller uniform divisor can stay within the limit.

The function:

* Preserves the total count and sum of the histogram.
* Only works on histogram metrics (no-op for other metric types).
* Makes no changes if:
  * The explicit boundary is not found when using `remove_explicit_bound`.
  * The histogram is already at or below the requested bucket limit when using `limit_buckets`.
  * The histogram is empty.
  * The histogram structure is invalid (mismatched bounds and counts, or unordered bounds when using `limit_buckets`).

**NOTE:** The `limit_buckets` method reduces histogram resolution and may affect percentile or quantile accuracy. Use only when you are confident that reducing bucket detail is acceptable for downstream consumers.

Examples:

```yaml theme={null}

- merge_histogram_buckets(0.5) where metric.name == "http_request_duration"

# Given a histogram with:
# bounds: [0.1, 0.5, 1.0]
# counts: [5, 8, 3, 1]
#
# After merging at 0.5:
# bounds: [0.1, 1.0]
# counts: [5, 11, 1]

# Limit histograms to at most 5 buckets
- merge_histogram_buckets(5, method="limit_buckets") where metric.name == "http_request_duration"

# Given a histogram with:
# bounds: [0.1, 0.2, 0.5, 1.0, 2.5, 5.0, 10.0, 30.0]
# counts: [80, 4, 6, 120, 3, 2, 40, 10, 1]
#
# After limiting to 5 buckets, one compaction pass merges bucket pairs:
# bounds: [0.2, 1.0, 5.0, 30.0]
# counts: [84, 126, 5, 50, 1]
```

### ParseCEF

`ParseCEF(target)`

The `ParseCEF` function returns a `pcommon.Map` that is the result of parsing the `target` string as a [Common Event Format (CEF)](https://www.microfocus.com/documentation/arcsight/arcsight-smartconnectors-8.4/cef-implementation-standard/Content/CEF/Chapter%201%20What%20is%20CEF.htm) message.

`target` is a Getter that returns a string. If the returned string is empty, or cannot be parsed as CEF, an error will be returned.

`ParseCEF` is tolerant of an optional syslog header preceding the `CEF:` token; parsing begins at the first occurrence of `CEF:` in the input.

The returned map has the following top-level fields:

* `cef.version` — the CEF version (the integer following `CEF:`).
* `cef.device_vendor`, `cef.device_product`, `cef.device_version`, `cef.device_event_class_id`, `cef.name`, `cef.severity` — the six CEF header fields.
* `cef.extensions` — a map of the parsed key/value extension pairs.

Within the header fields, the escape sequences `\|` (pipe) and `\\` (backslash) are unescaped. Within extension values, the escape sequences `\\` (backslash), `\=` (equals), `\n` (newline), and `\r` (carriage return) are unescaped.

Extension parsing uses the position of the next `key=` token as the end of the current value, so values may contain spaces. All extension values are returned as strings.

Examples:

* `ParseCEF(body)`

* `ParseCEF("CEF:0|Security|threatmanager|1.0|100|worm successfully stopped|10|src=10.0.0.1 dst=2.1.2.2 spt=1232")`

### ParseCLF

`ParseCLF(target, Optional[format])`

The `ParseCLF` function returns a `pcommon.Map` that is the result of parsing the `target` string as a [Common Log Format (CLF)](https://www.w3.org/Daemon/User/Config/Logging.html#common-logfile-format) HTTP access log entry.

`target` is a Getter that returns a string. If the returned string is empty, or cannot be parsed in the selected format, an error will be returned.

`format` is an optional string that selects the log format to parse. Valid values are:

* `"clf"` (default) — the strict Common Log Format:

  ```
  remotehost rfc931 auth_user [date] "request" status bytes
  ```

* `"combined"` — the NCSA Combined Log Format used by default in many Apache and nginx configurations, which is CLF with the quoted referer and user-agent appended:

  ```
  remotehost rfc931 auth_user [date] "request" status bytes "referer" "user-agent"
  ```

Quoted fields (`request`, `referer`, `user-agent`) may contain backslash escape sequences as produced by Apache (`\"`, `\\`, `\xhh`, and C-style control escapes such as `\n` and `\t` — see the [mod\_log\_config format notes](https://httpd.apache.org/docs/current/mod/mod_log_config.html#format-notes)) and nginx (`\xhh`). These sequences are unescaped in the returned values.

The returned map has the following fields:

* `clf.remote_host` — the client's DNS name or IP address.
* `clf.rfc931` — the remote logname of the user (CLF uses `-` when unknown).
* `clf.auth_user` — the authenticated user (CLF uses `-` when unknown).
* `clf.timestamp` — the contents of the bracketed date field, preserved as a string.
* `clf.request` — the raw request line as sent by the client.
* `clf.method`, `clf.request_uri`, `clf.protocol` — the parsed components of the request line, only set when the request line is well-formed.
* `clf.status` — the HTTP status code as an integer.
* `clf.bytes` — the content-length of the response as an integer. Omitted when CLF reports `-` (e.g. on a 304 response).
* `clf.referer`, `clf.user_agent` — the referer and user-agent strings, only set when `format` is `"combined"`.

Examples:

* `ParseCLF(body)`
* `ParseCLF("127.0.0.1 - frank [10/Oct/2000:13:55:36 -0700] \"GET /apache_pb.gif HTTP/1.0\" 200 2326")`
* `ParseCLF(body, "combined")`

### ParseELF

`ParseELF(target)`

The `ParseELF` function returns a `pcommon.Map` that is the result of parsing the `target` string as a [W3C Extended Log Format (ELF)](https://www.w3.org/TR/WD-logfile.html) log block.

`target` is a Getter that returns a string containing a complete ELF log block (one or more directive lines followed by data lines). If the string is empty or does not contain a valid `#Version` directive, an error is returned.

**Intended usage:** `ParseELF` is designed for pipelines where the full ELF block — header directives (`#Version`, `#Fields`, etc.) and data lines — is available as a single string. This is the case when using the [filelog receiver](../../receiver/filelogreceiver/README.md) with a multiline configuration that groups an entire ELF file (or rotated segment) into one log record body, or when the entire log content is read from a single field. In a line-by-line streaming pipeline where `#Fields` arrives only once at file open, each individual data line does not carry its own header; for that pattern you would need to prepend the known header directives to each data line before passing the string to `ParseELF`.

The returned map contains the following keys:

* `elf.version` — value of the `#Version` directive (required).
* `elf.software` — value of `#Software` (omitted if not present).
* `elf.date` — value of `#Date` (omitted if not present).
* `elf.start_date` — value of `#Start-Date` (omitted if not present).
* `elf.end_date` — value of `#End-Date` (omitted if not present).
* `elf.remark` — value of `#Remark` (omitted if not present).
* `elf.fields` — string slice of field names from the last `#Fields` directive.
* `elf.entries` — slice of maps, one per data line, keyed by field name. Missing values are represented as `"-"`.

Multiple `#Fields` directives within a single block are supported; each directive applies to subsequent data lines until the next `#Fields` directive is encountered. Double-quoted field values (as produced by Microsoft IIS) are handled correctly.

Examples:

* `ParseELF(body)`

* `ParseELF("#Version: 1.0\n#Fields: time cs-method cs-uri\n00:34:23 GET /foo/bar.html")`

### ParseLEEF

`ParseLEEF(target)`

The `ParseLEEF` function returns a `pcommon.Map` that is the result of parsing the `target` string as a [Log Event Extended Format (LEEF)](https://www.ibm.com/docs/en/dsm?topic=overview-leef-event-components) message.

`target` is a Getter that returns a string. If the returned string is empty, or cannot be parsed as LEEF, an error will be returned.

`ParseLEEF` can parse both LEEF 1.0 and LEEF 2.0 messages. The function is tolerant of an optional syslog header preceding the `LEEF:` token; parsing begins at the first occurrence of `LEEF:` in the input, so a literal `LEEF:` appearing in a syslog header ahead of the real header would be misinterpreted.

The returned map has the following top-level fields:

* `leef.version` — the LEEF version (`"1.0"` or `"2.0"`).
* `leef.vendor`, `leef.product.name`, `leef.product.version`, `leef.event.id` — the LEEF header fields.
* `leef.attributes` — a map of the parsed key/value attribute pairs.

For LEEF 1.0 the attribute delimiter is always a tab. For LEEF 2.0 the delimiter is taken from the header and must be either a single character or a `0x`-prefixed hex value decoding to a single byte (e.g. `0x09` for tab). An empty delimiter field defaults to tab. The delimiter field is also optional: if the position normally occupied by the delimiter looks like the start of an attribute (i.e. contains `=`), it is treated as the first attribute and the delimiter defaults to tab.

Attribute parsing is lenient: pairs without an `=` separator or with an empty key are silently skipped, and when the same key appears more than once the last occurrence wins. Whitespace within keys and values is preserved verbatim, since the LEEF spec defines a value as everything up to the delimiter.

All attribute values are returned as strings. LEEF defines a set of [predefined event attributes](https://www.ibm.com/docs/en/dsm?topic=overview-predefined-leef-event-attributes) (e.g. `src`, `dst`, `srcPort`, `usrName`, `devTime`) with expected types such as Integer, IPv4/IPv6, and Time.

Examples:

* `ParseLEEF(body)`

* `ParseLEEF("LEEF:1.0|Microsoft|MSExchange|4.0 SP1|15345|src=10.50.1.1\tdst=2.10.20.20\tsev=5")`

* `ParseLEEF("LEEF:2.0|Lancope|StealthWatch|1.0|41|^|src=10.0.1.8^dst=10.0.0.5^sev=5")`

### set\_semconv\_span\_name

`set_semconv_span_name(semconvVersion, Optional[originalSpanNameAttribute])`

The `set_semconv_span_name()` function overwrites a span name using the OpenTelemetry semantic conventions for [HTTP](https://opentelemetry.io/docs/specs/semconv/http/http-spans/), [RPC](https://opentelemetry.io/docs/specs/semconv/rpc/rpc-spans/), [messaging](https://opentelemetry.io/docs/specs/semconv/messaging/messaging-spans/), and [database](https://opentelemetry.io/docs/specs/semconv/database/) spans. In other cases, the original `span.name` remains unchanged.

The primary use case of the `set_semconv_span_name()` function is to address high-cardinality issues in span metrics when `span.name` doesn't comply with the OpenTelemetry requirement that span names be low cardinality such as `GET /product/12345`, `GET/product?id=12345`, or `SELECT * FROM product WHERE id=12345`.

Parameters:

* `semconvVersion` is the version of the Semantic Conventions used to generate the `span.name`. A set of older/deprecated semconv attributes are also supported for backward compatibility (see the table below). Versions `1.37.0` to `1.43.0` are supported.
* `originalSpanNameAttribute` is the optional name of the attribute used to copy the original `span.name` if different from the name derived from semantic conventions.

Sanitization examples:

* Span with high-cardinality name but recommended semantic convention attributes
  * Incoming span:
    ```
    span.name: GET /api/v1/users/123 # /!\ high cardinality
    span.kind: server
    span.attributes
       http.request.method: GET
       http.route: /api/v1/users/{id}
       url.path: /api/v1/users/123
    ```
  * Span name after applying `set_semconv_span_name("1.43.0")`: `GET /api/v1/users/{id}`
  * No loss of information on `span.name` occurs because the recommended attribute `http.route` is present.
* Span with high-cardinality name lacking recommended semantic convention attribute `http.route`
  * Incoming span:
    ```
    span.name: GET /api/v1/users/123 # /!\ high cardinality
    span.kind: server
    span.attributes
       http.request.method: GET
       url.path: /api/v1/users/123
    ```
  * Span name after applying `set_semconv_span_name("1.43.0")`: `GET`
  * Loss of information on `span.name` occurs because the recommended attribute `http.route` is missing.
    Note that this loss of information is mitigated if the instrumentation produced attributes that contain the URL path like `url.path` or `url.full`.
* Compliant span name is unchanged
  * Incoming span:
    ```
    span.name: GET /api/v1/users/{id}
    span.kind: server
    span.attributes
       http.request.method: GET
       http.route: /api/v1/users/{id}
       url.path: /api/v1/users/123
    ```
  * Span name after applying `set_semconv_span_name("1.43.0")`: `GET /api/v1/users/{id}`

Backward compatibility: `set_semconv_span_name` will also read the following deprecated attributes, mapping each to its current equivalent:

| v1.43.0 Attribute | Older attribute |
| - | - |
| `http.request.method` | `http.method` |
| `rpc.method` | `rpc.grpc.method` |
| `rpc.service` | `rpc.grpc.service` |
| `rpc.system.name` | `rpc.system` |
| `db.system.name` | `db.system` |
| `db.operation.name` | `db.operation` |
| `db.namespace` | `db.name` |
| `messaging.operation.name` | `messaging.operation` |
| `messaging.destination.name` | `messaging.destination` |

Examples:

* `set_semconv_span_name("1.43.0")`

* `set_semconv_span_name("1.43.0", "original_span_name")`

## Examples

### Perform transformation if field does not exist

Set attribute `test` to `"pass"` if the attribute `test` does not exist:

```yaml theme={null}
transform:
  error_mode: ignore
  trace_statements:
    # accessing a map with a key that does not exist will return nil. 
    - set(span.attributes["test"], "pass") where span.attributes["test"] == nil
```

### Rename attribute

There are 2 ways to rename an attribute key:

You can either set a new attribute and delete the old:

```yaml theme={null}
transform:
  error_mode: ignore
  trace_statements:
    - set(resource.attributes["namespace"], resource.attributes["k8s.namespace.name"])
    - delete_key(resource.attributes, "k8s.namespace.name") 
```

Or you can update the key using regex:

```yaml theme={null}
transform:
  error_mode: ignore
  trace_statements:
    - replace_all_patterns(resource.attributes, "key", "k8s\\.namespace\\.name", "namespace")
```

### Move field to attribute

Set attribute `body` to the value of the log body:

```yaml theme={null}
transform:
  error_mode: ignore
  log_statements:
    - set(log.attributes["body"], log.body)
```

### Combine two attributes

Set attribute `test` to the value of attributes `"foo"` and `"bar"` combined.

```yaml theme={null}
transform:
  error_mode: ignore
  trace_statements:
    # Use Concat function to combine any number of string, separated by a delimiter.
    - set(resource.attributes["test"], Concat([resource.attributes["foo"], resource.attributes["bar"]], " "))
```

### Parsing JSON logs

Given the following json body

```json theme={null}
{
  "name": "log",
  "attr1": "foo",
  "attr2": "bar",
  "nested": {
    "attr3": "example"
  }
}
```

add specific fields as attributes on the log:

```yaml theme={null}
transform:
  log_statements:
    - statements:
        # Parse body as JSON and merge the resulting map with the cache map, ignoring non-json bodies.
        # cache is a field exposed by OTTL that is a temporary storage place for complex operations.
        - merge_maps(log.cache, ParseJSON(log.body), "upsert") where IsMatch(log.body, "^\\{") 
          
        # Set attributes using the values merged into cache.
        # If the attribute doesn't exist in cache then nothing happens.
        - set(log.attributes["attr1"], log.cache["attr1"])
        - set(log.attributes["attr2"], log.cache["attr2"])
        
        # To access nested maps you can chain index ([]) operations.
        # If nested or attr3 do not exist in cache then nothing happens.
        - set(log.attributes["nested.attr3"], log.cache["nested"]["attr3"])
```

### Override context statements error mode

```yaml theme={null}
transform:
  # default error mode applied to all context statements
  error_mode: propagate
  log_statements:
    # overrides the default error mode for these statements
    - error_mode: ignore
      statements:
        - merge_maps(log.cache, ParseJSON(log.body), "upsert") where IsMatch(log.body, "^\\{")
        - set(log.attributes["attr1"], log.cache["attr1"])

    # uses the default error mode
    - statements:
        - set(log.attributes["namespace"], log.attributes["k8s.namespace.name"])
```

### Get Severity of an Unstructured Log Body

Given the following unstructured log body

```txt theme={null}
[2023-09-22 07:38:22,570] INFO [Something]: some interesting log
```

You can find the severity using IsMatch:

```yaml theme={null}
transform:
  error_mode: ignore
  log_statements:
    - set(log.severity_number, SEVERITY_NUMBER_INFO) where IsString(log.body) and IsMatch(log.body, "\\sINFO\\s")
    - set(log.severity_number, SEVERITY_NUMBER_WARN) where IsString(log.body) and IsMatch(log.body, "\\sWARN\\s")
    - set(log.severity_number, SEVERITY_NUMBER_ERROR) where IsString(log.body) and IsMatch(log.body, "\\sERROR\\s")
```

## Copy attributes matching regular expression to a separate location

If you want to move resource attributes, which keys are matching the regular expression `pod_labels_.*` to a new attribute
location `kubernetes.labels`, use the following configuration:

```yaml theme={null}
transform:
  error_mode: ignore
  trace_statements:
    - statements:
        - set(resource.cache["attrs"], resource.attributes)
        - keep_matching_keys(resource.cache["attrs"], "pod_labels_.*")
        - set(resource.attributes["kubernetes.labels"], resource.cache["attrs"])
```

The configuration can be used also with `delete_matching_keys()` to copy the attributes that do not match the regular expression.

## Troubleshooting

When using OTTL you can enable debug logging in the collector to print out useful information,
such as the current Statement and the current TransformContext, to help you troubleshoot
why a statement is not behaving as you expect. This feature is very verbose, but provides you an accurate
view into how OTTL views the underlying data.

```yaml theme={null}
receivers:
  file_log:
    start_at: beginning
    include: [ test.log ]

processors:
  transform:
    error_mode: ignore
    log_statements:
      - set(resource.attributes["test"], "pass")
      - set(scope.attributes["test"], ["pass"])
      - set(log.attributes["test"], true)
          

exporters:
  debug:

service:
  telemetry:
    logs:
      level: debug
  pipelines:
    logs:
      receivers:
        - file_log
      processors:
        - transform
      exporters:
        - debug
```

```
2025-02-13T13:01:07.590-0700    debug   ottl@v0.119.0/parser.go:356     initial TransformContext before executing StatementSequence     {"otelcol.component.id": "transform", "otelcol.component.kind": "Processor", "otelcol.pipeline.id": "logs", "otelcol.signal": "logs", "TransformContext": {"resource": {"attributes": {}, "dropped_attribute_count": 0}, "cache": {}}}
2025-02-13T13:01:07.591-0700    debug   ottl@v0.119.0/parser.go:35      TransformContext after statement execution      {"otelcol.component.id": "transform", "otelcol.component.kind": "Processor", "otelcol.pipeline.id": "logs", "otelcol.signal": "logs", "statement": "set(resource.attributes[\"test\"], \"pass\")", "condition matched": true, "TransformContext": {"resource": {"attributes": {"test": "pass"}, "dropped_attribute_count": 0}, "cache": {}}}
2025-02-13T13:01:07.593-0700    debug   ottl@v0.119.0/parser.go:356     initial TransformContext before executing StatementSequence     {"otelcol.component.id": "transform", "otelcol.component.kind": "Processor", "otelcol.pipeline.id": "logs", "otelcol.signal": "logs", "TransformContext": {"resource": {"attributes": {"test": "pass"}, "dropped_attribute_count": 0}, "scope": {"attributes": {}, "dropped_attribute_count": 0, "name": "", "version": ""}, "cache": {}}}
2025-02-13T13:01:07.594-0700    debug   ottl@v0.119.0/parser.go:35      TransformContext after statement execution      {"otelcol.component.id": "transform", "otelcol.component.kind": "Processor", "otelcol.pipeline.id": "logs", "otelcol.signal": "logs", "statement": "set(scope.attributes[\"test\"], [\"pass\"])", "condition matched": true, "TransformContext": {"resource": {"attributes": {"test": "pass"}, "dropped_attribute_count": 0}, "scope": {"attributes": {"test": ["pass"]}, "dropped_attribute_count": 0, "name": "", "version": ""}, "cache": {}}}
2025-02-13T13:01:07.594-0700    debug   ottl@v0.119.0/parser.go:356     initial TransformContext before executing StatementSequence     {"otelcol.component.id": "transform", "otelcol.component.kind": "Processor", "otelcol.pipeline.id": "logs", "otelcol.signal": "logs", "TransformContext": {"resource": {"attributes": {"test": "pass"}, "dropped_attribute_count": 0}, "scope": {"attributes": {"test": ["pass"]}, "dropped_attribute_count": 0, "name": "", "version": ""}, "log_record": {"attributes": {"log.file.name": "test.log"}, "body": "test", "dropped_attribute_count": 0, "flags": 0, "observed_time_unix_nano": 1739476867483160000, "severity_number": 0, "severity_text": "", "span_id": "0000000000000000", "time_unix_nano": 0, "trace_id": "00000000000000000000000000000000"}, "cache": {}}}
2025-02-13T13:01:07.594-0700    debug   ottl@v0.119.0/parser.go:35      TransformContext after statement execution      {"otelcol.component.id": "transform", "otelcol.component.kind": "Processor", "otelcol.pipeline.id": "logs", "otelcol.signal": "logs", "statement": "set(log.attributes[\"test\"], true)", "condition matched": true, "TransformContext": {"resource": {"attributes": {"test": "pass"}, "dropped_attribute_count": 0}, "scope": {"attributes": {"test": ["pass"]}, "dropped_attribute_count": 0, "name": "", "version": ""}, "log_record": {"attributes": {"log.file.name": "test.log", "test": true}, "body": "test", "dropped_attribute_count": 0, "flags": 0, "observed_time_unix_nano": 1739476867483160000, "severity_number": 0, "severity_text": "", "span_id": "0000000000000000", "time_unix_nano": 0, "trace_id": "00000000000000000000000000000000"}, "cache": {}}}
2025-02-13T13:01:07.594-0700    info    Logs    {"otelcol.component.id": "debug", "otelcol.component.kind": "Exporter", "otelcol.signal": "logs", "resource logs": 1, "log records": 1}
```

## Contributing

See [CONTRIBUTING.md](https://github.com/open-telemetry/opentelemetry-collector-contrib/blob/main/processor/transformprocessor/CONTRIBUTING.md).

## Warnings

The Transform Processor uses the [OpenTelemetry Transformation Language](https://github.com/open-telemetry/opentelemetry-collector-contrib/blob/main/pkg/ottl/README.md) (OTTL) which allows users to modify all aspects of their telemetry. Some specific risks are listed below, but this is not an exhaustive list. In general, understand your data before using the Transform Processor.

* [Unsound Transformations](https://github.com/open-telemetry/opentelemetry-collector/blob/main/docs/standard-warnings.md#unsound-transformations): Several Metric-only functions allow you to transform one metric data type to another or create new metrics from an existing metrics.  Transformations between metric data types are not defined in the [metrics data model](https://github.com/open-telemetry/opentelemetry-specification/blob/main//specification/metrics/data-model.md).  These functions have the expectation that you understand the incoming data and know that it can be meaningfully converted to a new metric data type or can meaningfully be used to create new metrics.
  * Although the OTTL allows the `set` function to be used with `metric.data_type`, its implementation in the Transform Processor is NOOP.  To modify a data type you must use a function specific to that purpose.
* [Identity Conflict](https://github.com/open-telemetry/opentelemetry-collector/blob/main/docs/standard-warnings.md#identity-conflict): Transformation of metrics have the potential to affect the identity of a metric leading to an Identity Crisis. Be especially cautious when transforming metric name and when reducing/changing existing attributes.  Adding new attributes is safe.
* [Orphaned Telemetry](https://github.com/open-telemetry/opentelemetry-collector/blob/main/docs/standard-warnings.md#orphaned-telemetry): The processor allows you to modify `span_id`, `trace_id`, and `parent_span_id` for traces and `span_id`, and `trace_id` logs.  Modifying these fields could lead to orphaned spans or logs.
* Other: Statements that modify a higher context (`resource` or `scope`) using values from a lower context (for example, `set(resource.attributes["to"], log.attributes["from"])`) may produce unexpected results. The statement runs once per span, metric, or log record, and every record that shares the same resource or scope writes to the same shared value, so the last record processed determines the final value. For logs, the [`flatten_data`](#transformflattenlogs) option can be used to avoid this. See [#32080](https://github.com/open-telemetry/opentelemetry-collector-contrib/issues/32080) for more details.

## Feature Gate

### `transform.flatten.logs`

The `transform.flatten.logs` [feature gate](https://github.com/open-telemetry/opentelemetry-collector/blob/main/featuregate/README.md#collector-feature-gates) enables the `flatten_data` configuration option (default `false`). With `flatten_data: true`, the processor provides each log record with a distinct copy of its resource and scope. Then, after applying all transformations, the log records are regrouped by resource and scope.

This option is useful when applying transformations which alter the resource or scope. e.g. `set(resource.attributes["to"], log.attributes["from"])`, which may otherwise result in unexpected behavior. Using this option typically incurs a performance penalty as the processor must compute many hashes and create copies of resource and scope information for every log record.

The feature is currently only available for log processing.

#### Example Usage

`config.yaml`:

```yaml theme={null}
transform:
  flatten_data: true
  log_statements:
    - set(resource.attributes["to"], log.attributes["from"])
```

Run collector: `./otelcol --config config.yaml --feature-gates=transform.flatten.logs`

### `ottl.set.allowNil`

The `ottl.set.allowNil` [feature gate](https://github.com/open-telemetry/opentelemetry-collector-contrib/blob/main/pkg/ottl/documentation.md) changes the behavior of the OTTL `set` function when a `nil` value is evaluated. When enabled, `set` will pass the `nil` value directly to the target. Depending on the target, this may result in an error or an empty value. See the [OTTL Documentation](https://github.com/open-telemetry/opentelemetry-collector-contrib/tree/main/pkg/ottl) for full details and migration instructions.

## Compatibility

The Transform Processor operates only on the telemetry flowing through the pipeline and has no
external service or system dependencies. Its behavior is defined entirely by the OTTL functions,
paths, and enums it exposes, which are versioned together with the processor in each OpenTelemetry
Collector release. There are no external dependency versions to track for compatibility.

Configuration syntax and OTTL semantics follow the [OpenTelemetry Collector's versioning and
stability policies](https://github.com/open-telemetry/opentelemetry-collector/blob/main/docs/component-stability.md).
Breaking changes to configuration or OTTL behavior are gated behind [feature gates](#feature-gate)
and announced in the [changelog](https://github.com/open-telemetry/opentelemetry-collector-contrib/blob/main/CHANGELOG.md).

## Performance and scaling

The Transform Processor is stateless across batches: each incoming batch is processed independently,
so the processor requires no coordination and can be scaled horizontally by running additional
Collector replicas. The `shared_cache` option and OTTL's `cache` only hold state for the duration of
a single execution and are never shared across batches or replicas.

CPU and memory cost scale with the number and complexity of the statements you configure and with the
volume of telemetry flowing through the pipeline. To keep the processor efficient:

* Use conditions (statement-level `where` clauses or advanced [global conditions](#advanced-config))
  to skip statements for telemetry they don't apply to.
* Prefer fewer, more targeted statements over broad transformations applied to all telemetry.
* The [`flatten_data`](#transformflattenlogs) option incurs a performance penalty because the
  processor must compute hashes and copy resource and scope information for every log record.

The transform processor is tested as part of the project's load tests, with the results publicly
available on the benchmarks
[page](https://open-telemetry.github.io/opentelemetry-collector-contrib/benchmarks/loadtests).
There you can find the CPU and memory usage of the processor executing a representative set of OTTL
statements against each signal at 10,000 items/second:

* Traces: [CPU](https://open-telemetry.github.io/opentelemetry-collector-contrib/benchmarks/loadtests/#transformprocessortraces-cpu-percentage) and [memory](https://open-telemetry.github.io/opentelemetry-collector-contrib/benchmarks/loadtests/#transformprocessortraces-ram-mib)
* Metrics: [CPU](https://open-telemetry.github.io/opentelemetry-collector-contrib/benchmarks/loadtests/#transformprocessormetrics-cpu-percentage) and [memory](https://open-telemetry.github.io/opentelemetry-collector-contrib/benchmarks/loadtests/#transformprocessormetrics-ram-mib)
* Logs: [CPU](https://open-telemetry.github.io/opentelemetry-collector-contrib/benchmarks/loadtests/#transformprocessorlogs-cpu-percentage) and [memory](https://open-telemetry.github.io/opentelemetry-collector-contrib/benchmarks/loadtests/#transformprocessorlogs-ram-mib)

Refer to the [test](https://github.com/open-telemetry/opentelemetry-collector-contrib/blob/main/testbed/tests/transform_processor_test.go) for more information about the setup.

Each function that is unique to the transform processor also has a Go benchmark covering its
execution. Run them with:

```sh theme={null}
make benchmark-transform
```

## Self-observability

The Transform Processor does not emit any component-specific metrics or spans. It relies on
the standard telemetry that the Collector produces for every processor, which is described in the
Collector's [internal telemetry](https://opentelemetry.io/docs/collector/internal-telemetry/)
documentation.

### Metrics

The following metrics are emitted for the processor:

* `otelcol_processor_incoming_items`: the number of items (spans, metric data points, or log records)
  passed to the processor.
* `otelcol_processor_outgoing_items`: the number of items emitted by the processor.
* `otelcol_processor_internal_duration`: a histogram of the time taken to process a batch of
  telemetry through the processor.

These metrics carry the `otel.signal` and `processor` attributes, so you can observe throughput per
signal and per configured transform processor instance.

### Logs

The amount of detail the processor logs is controlled by the Collector's
[`service::telemetry::logs::level`](https://opentelemetry.io/docs/collector/internal-telemetry/#configure-internal-logs)
setting:

* At `info` (the default), the processor logs no per-statement information. Errors are logged
  according to the configured [`error_mode`](#general-config): `ignore` logs each failed statement or
  condition at `warn` level, `silent` suppresses them entirely, and `propagate` returns the error up
  the pipeline (where the payload is dropped and the failure is surfaced by the pipeline rather than
  by the processor).
* At `debug`, OTTL additionally logs the details it uses to evaluate your statements: the OTTL context
  it infers for each statement group, the `TransformContext` before a statement sequence runs and
  after every individual statement executes, and the result of each condition evaluation. This is very
  verbose but is the most useful level for debugging why a statement is or isn't behaving as expected,
  since it shows exactly how OTTL sees the underlying data. See [Troubleshooting](#troubleshooting) for
  an example of this output.

## Configuration

### Example Configuration

```yaml theme={null}
transform:
  trace_statements:
    - context: span
      statements:
        - set(name, "bear") where attributes["http.path"] == "/animal"
        - keep_keys(attributes, ["http.method", "http.path"])
    - context: resource
      statements:
        - set(attributes["name"], "bear")
  metric_statements:
    - context: datapoint
      statements:
        - set(metric.name, "bear") where attributes["http.path"] == "/animal"
        - keep_keys(attributes, ["http.method", "http.path"])
    - context: resource
      statements:
        - set(attributes["name"], "bear")
  log_statements:
    - context: log
      statements:
        - set(body, "bear") where attributes["http.path"] == "/animal"
        - keep_keys(attributes, ["http.method", "http.path"])
    - context: resource
      statements:
        - set(attributes["name"], "bear")
  profile_statements:
    - context: profile
      statements:
        - set(original_payload_format, "bear") where original_payload_format == "/animal"
    - context: resource
      statements:
        - set(attributes["name"], "bear")

transform/with_conditions:
  trace_statements:
    - context: span
      conditions:
        - attributes["http.path"] == "/animal"
      statements:
        - set(name, "bear")
  metric_statements:
    - context: datapoint
      conditions:
        - attributes["http.path"] == "/animal"
      statements:
        - set(metric.name, "bear")
  log_statements:
    - context: log
      conditions:
        - attributes["http.path"] == "/animal"
      statements:
        - set(body, "bear")     
  profile_statements:
    - context: profile
      conditions:
        - original_payload_format == "/animal"
      statements:
        - set(original_payload_format, "bear")

transform/ignore_errors:
  error_mode: ignore
  trace_statements:
    - context: resource
      statements:
        - set(attributes["name"], "bear")

transform/bad_syntax_log:
  log_statements:
    - context: log
      statements:
        - set(body, "bear" where attributes["http.path"] == "/animal"
        - keep_keys(attributes, ["http.method", "http.path"])

transform/bad_syntax_metric:
  metric_statements:
    - context: datapoint
      statements:
        - set(name, "bear" where attributes["http.path"] == "/animal"
        - keep_keys(attributes, ["http.method", "http.path"])

transform/bad_syntax_trace:
  trace_statements:
    - context: span
      statements:
        - set(name, "bear" where attributes["http.path"] == "/animal"
        - keep_keys(attributes, ["http.method", "http.path"])

transform/bad_syntax_profile:
  log_statements:
    - context: profile
      statements:
        - set(original_payload_format, "bear" where attributes["http.path"] == "/animal"
        - keep_keys(attributes, ["http.method", "http.path"])

transform/bad_syntax_multi_signal:
  trace_statements:
    - context: span
      statements:
        - set(name, "bear" where attributes["http.path"] == "/animal"
        - keep_keys(attributes, ["http.method", "http.path"])
  metric_statements:
    - context: datapoint
      statements:
        - set(name, "bear" attributes["http.path"] == "/animal"
        - keep_keys(attributes, ["http.method", "http.path"])
  log_statements:
    - context: log
      statements:
        - set(body, "bear" none["http.path"] == "/animal"
        - keep_keys(attributes, ["http.method", "http.path"])
  profile_statements:
    - context: profile
      statements:
        - set(original_payload_format, "bear" none["http.path"] == "/animal"
        - keep_keys(attributes, ["http.method", "http.path"])

transform/unknown_function_log:
  log_statements:
    - context: log
      statements:
        - set(body, "bear") where attributes["http.path"] == "/animal"
        - not_a_function(attributes, ["http.method", "http.path"])

transform/unknown_function_metric:
  metric_statements:
    - context: datapoint
      statements:
        - set(metric.name, "bear") where attributes["http.path"] == "/animal"
        - not_a_function(attributes, ["http.method", "http.path"])

transform/unknown_function_trace:
  trace_statements:
    - context: span
      statements:
        - set(name, "bear") where attributes["http.path"] == "/animal"
        - not_a_function(attributes, ["http.method", "http.path"])

transform/unknown_function_profile:
  log_statements:
    - context: profile
      statements:
        - set(original_payload_format, "bear") where attributes["http.path"] == "/animal"
        - not_a_function(attributes, ["http.method", "http.path"])

transform/unknown_context:
  trace_statements:
    - context: test
      statements:
        - set(name, "bear") where attributes["http.path"] == "/animal"

transform/unknown_error_mode:
  error_mode: test

transform/structured_configuration_with_path_context:
  trace_statements:
    - context: span
      statements:
        - set(span.name, "bear") where span.attributes["http.path"] == "/animal"
  metric_statements:
    - context: metric
      statements:
        - set(metric.name, "bear") where resource.attributes["http.path"] == "/animal"
  log_statements:
    - context: log
      statements:
        - set(log.body, "bear") where log.attributes["http.path"] == "/animal"
  profile_statements:
    - context: profile
      statements:
        - set(profile.original_payload_format, "bear") where profile.original_payload_format == "/animal"

transform/structured_configuration_with_inferred_context:
  trace_statements:
    - statements:
      - set(span.name, "bear") where span.attributes["http.path"] == "/animal"
      - set(resource.attributes["name"], "bear")
  metric_statements:
    - statements:
      - set(metric.name, "bear") where resource.attributes["http.path"] == "/animal"
      - set(resource.attributes["name"], "bear")
  log_statements:
    - statements:
      - set(log.body, "bear") where log.attributes["http.path"] == "/animal"
      - set(resource.attributes["name"], "bear")
  profile_statements:
    - statements:
        - set(profile.original_payload_format, "bear") where profile.original_payload_format == "/animal"
        - set(resource.attributes["name"], "bear")

transform/flat_configuration:
  trace_statements:
    - set(span.name, "bear") where span.attributes["http.path"] == "/animal"
    - set(resource.attributes["name"], "bear")
  metric_statements:
    - set(metric.name, "bear") where resource.attributes["http.path"] == "/animal"
    - set(resource.attributes["name"], "bear")
  log_statements:
    - set(log.body, "bear") where log.attributes["http.path"] == "/animal"
    - set(resource.attributes["name"], "bear")
  profile_statements:
    - set(profile.original_payload_format, "bear") where profile.original_payload_format == "/animal"
    - set(resource.attributes["name"], "bear")

transform/mixed_configuration_styles:
  trace_statements:
    - set(span.name, "bear") where span.attributes["http.path"] == "/animal"
    - context: span
      statements:
        - set(attributes["name"], "bear")
        - keep_keys(attributes, ["http.method", "http.path"])

transform/with_shared_cache_key:
  trace_statements:
    - statements:
        - set(resource.attributes["name"], "propagate")
  metric_statements:
    - statements:
        - set(resource.attributes["name"], "silent")
      shared_cache: true

transform/context_statements_error_mode:
  error_mode: ignore
  trace_statements:
    - error_mode: propagate
      statements:
        - set(resource.attributes["name"], "propagate")
    - statements:
        - set(resource.attributes["name"], "ignore")
  metric_statements:
    - error_mode: silent
      statements:
        - set(resource.attributes["name"], "silent")
    - statements:
        - set(resource.attributes["name"], "ignore")
  log_statements:
    - error_mode: propagate
      statements:
        - set(resource.attributes["name"], "propagate")
    - statements:
        - set(resource.attributes["name"], "ignore")
  profile_statements:
    - error_mode: propagate
      statements:
        - set(resource.attributes["name"], "propagate")
    - statements:
        - set(resource.attributes["name"], "ignore")

transform/shared_cache:
  trace_statements:
    - shared_cache: true
      statements:
        - set(resource.attributes["name"], "bear")
    - shared_cache: true
      statements:
        - set(resource.attributes["copy"], resource.attributes["name"])
  metric_statements:
    - shared_cache: true
      statements:
        - set(resource.attributes["name"], "bear")
  log_statements:
    - shared_cache: true
      statements:
        - set(resource.attributes["name"], "bear")
```

***

*Last generated: 2026-10-05*