Sampling

Adding Sentry to your app gives you a great deal of very valuable information about errors and performance you wouldn't otherwise get. And lots of information is good -- as long as it's the right information, at a reasonable volume.

To send a representative sample of your errors to Sentry, set the sample_rate option in your SDK configuration to a number between 0 (0% of errors sent) and 1 (100% of errors sent). This is a static rate, which will apply equally to all errors. For example, to sample 25% of your errors:

Copied
import sentry_sdk

sentry_sdk.init(
    # ...

    sample_rate=0.25,
)

The error sample rate defaults to 1.0, meaning all errors are sent to Sentry.

To sample error events dynamically, set the error_sampler to a function that returns the desired sample rate for the event. The error_sampler takes two arguments, event and hint, which it uses to inform the sampling decision.

One potential use case for the error_sampler is to apply different sample rates for different exception types. For instance, if you would like to sample some exception called MyException at 50%, discard all events of another exception called MyIgnoredException, and sample all other exception types at 100%, you could use the following code when initializing the SDK:

Copied
import sentry_sdk

def my_sampler(_, hint):
    exception_sampler_values = {
        MyException: 0.5,
        MyIgnoredException: 0.0,  # or equivalently, False
    }

    try:
        return exception_sampler_values[hint["exc_info"][0]]
    except (IndexError, KeyError, TypeError):
        return 1.0  # or equivalently, True

sentry_sdk.init(
    # ...

    error_sampler=my_sampler,
)

We recommend sampling your transactions for two reasons:

  1. Capturing a single trace involves minimal overhead, but capturing traces for every page load or every API request may add an undesirable load to your system.
  2. Enabling sampling allows you to better manage the number of events sent to Sentry, so you can tailor your volume to your organization's needs.

Choose a sampling rate with the goal of finding a balance between performance and volume concerns with data accuracy. You don't want to collect too much data, but you want to collect sufficient data from which to draw meaningful conclusions. If you’re not sure what rate to choose, start with a low value and gradually increase it as you learn more about your traffic patterns and volume.

The Sentry SDKs have two configuration options to control the volume of transactions sent to Sentry, allowing you to take a representative sample:

  1. Uniform sample rate (traces_sample_rate):
    • Provides an even cross-section of transactions, no matter where in your app or under what circumstances they occur.
    • Uses default inheritance and precedence behavior
  2. Sampling function (traces_sampler) which:

By default, none of these options are set, meaning no transactions will be sent to Sentry. You must set either one of the options to start sending transactions.

To do this, set the traces_sample_rate option in your sentry_sdk.init() to a number between 0 and 1. With this option set, every transaction created will have that percentage chance of being sent to Sentry. (So, for example, if you set traces_sample_rate to 0.2, approximately 20% of your transactions will get recorded and sent.) That looks like this:

Copied
sentry_sdk.init(
    # ...

    # Set traces_sample_rate to 1.0 to capture 100%
    # of transactions for performance monitoring.
    # We recommend adjusting this value in production,
    traces_sample_rate=1.0,
)

To use the sampling function, set the traces_sampler option in your sentry-sdk.init() to a function that will accept a sampling_context dictionary and return a sample rate between 0 and 1. For example:

Copied
def traces_sampler(sampling_context):
    # Examine provided context data (including parent decision, if any)
    # along with anything in the global namespace to compute the sample rate
    # or sampling decision for this transaction

    if "...":
        # These are important - take a big sample
        return 0.5
    elif "...":
        # These are less important or happen much more frequently - only take 1%
        return 0.01
    elif "...":
        # These aren't something worth tracking - drop all transactions like this
        return 0
    else:
        # Default sample rate
        return 0.1

sentry_sdk.init(
    # ...

    traces_sampler=traces_sampler,
)

For convenience, the function can also return a boolean. Returning True is equivalent to returning 1, and will guarantee the transaction will be sent to Sentry. Returning False is equivalent to returning 0 and will guarantee the transaction will not be sent to Sentry.

The information contained in the sampling_context object passed to the traces_sampler when a transaction is created varies by platform and integration.

For Python-based SDKs, it includes at least the following:

Copied
{
  "transaction_context": {
    "name": <string>  # transaction title at creation time
    "op": <string>  # short description of transaction type, like "http.request"
  },
  "parent_sampled": <bool>  # if this transaction has a parent, its sampling decision
  ...  # custom context as passed to `start_transaction`
}

When using custom instrumentation to create a transaction, you can add data to the sampling_context by passing it as an optional second argument to start_transaction. This is useful if there's data to which you want the sampler to have access but which you don't want to attach to the transaction as tags or data, such as information that's sensitive or that’s too large to send with the transaction. For example:

Copied
sentry_sdk.start_transaction(
    # kwargs passed to Transaction constructor - will be recorded on transaction
    name="GET /search",
    op="search",
    data={
      "query_params": {
        "animal": "dog",
        "type": "very good"
      }
    },
    # `custom_sampling_context` - won't be recorded
    custom_sampling_context={
        # PII
        "user_id": "12312012",
        # too big to send
        "search_results": { ... }
    }
);

Whatever a transaction's sampling decision, that decision will be passed to its child spans and from there to any transactions they subsequently cause in other services.

(See Distributed Tracing for more about how that propagation is done.)

If the transaction currently being created is one of those subsequent transactions (in other words, if it has a parent transaction), the upstream (parent) sampling decision will be included in the sampling context data. Your traces_sampler can use this information to choose whether to inherit that decision. In most cases, inheritance is the right choice, to avoid breaking distributed traces. A broken trace will not include all your services.

Copied
def traces_sampler(sampling_context):
    # always inherit
    if sampling_context["parent_sampled"] is not None:
        return sampling_context["parent_sampled"]

    ...
    # rest of sampling logic here

If you're using a traces_sample_rate rather than a traces_sampler, the decision will always be inherited.

If you know at transaction creation time whether or not you want the transaction sent to Sentry, you also have the option of passing a sampling decision directly to the transaction constructor (note, not in the custom_sampling_context object). If you do that, the transaction won't be subject to the traces_sample_rate, nor will traces_sampler be run, so you can count on the decision that's passed not to be overwritten.

Copied
sentry_sdk.start_transaction(
  name="GET /search",
  sampled=True
);

There are multiple ways for a transaction to end up with a sampling decision.

  • Random sampling according to a static sample rate set in traces_sample_rate
  • Random sampling according to a sample function rate returned by traces_sampler
  • Absolute decision (100% chance or 0% chance) returned by traces_sampler
  • If the transaction has a parent, inheriting its parent's sampling decision
  • Absolute decision passed to start_transaction

When there's the potential for more than one of these to come into play, the following precedence rules apply:

  1. If a sampling decision is passed to start_transaction, that decision will be used, overriding everything else.
  2. If traces_sampler is defined, its decision will be used. It can choose to keep or ignore any parent sampling decision, use the sampling context data to make its own decision, or choose a sample rate for the transaction. We advise against overriding the parent sampling decision because it will break distributed traces)
  3. If traces_sampler is not defined, but there's a parent sampling decision, the parent sampling decision will be used.
  4. If traces_sampler is not defined and there's no parent sampling decision, traces_sample_rate will be used.
Help improve this content
Our documentation is open source and available on GitHub. Your contributions are welcome, whether fixing a typo (drat!) or suggesting an update ("yeah, this would be better").