NB5 Docs► User Guide► Advanced Topics▼ Labeling Results 🖺

👉 This provides an intro and overview of the built-in labeling system in NoSQLBench 5.21.

What are Labels?

NoSQLBench allows for your metrics and accompanying annotations to be organized in a self-consistent way with a set of labels and tags. The labeling facility is built-in to NoSQLBench -- Labeling is not something you add to existing tests results after using it. Labels are innate, self-consistent, and easily leveraged to apply appropriate metadata to your results so that they are well-defined and easily repurposed in other systems.

For example, when you name your activity, you are setting a label, even if you didn't know it already. You are setting a label of activity=myactivityname. This builds on top of the name everything strategy so that everything from your session down to your op level has a name and accompanying label. This system is implemented efficiently as well, with all labeling mechanisms being applied explicitly during initialization phases, rather than lazily during an activity.

Labels are defensive in nature. If you try to define another value for a label which has already been provided for a given runtime scope, an error will be thrown. This ensures that labeling semantics are protected and that users are informed when there would be a labeling conflict. This helps avoid structural or logical errors in their test setup.

Labels represent a implied hierarchic structure, and this is how they are assembled. The tree-like structure of components in the runtime allows for labels to be established from the trunk to the branches, with each label attaching to a specific component in the tree. However, when presented for a given component, they are simply a map, including all labels which are inherited from the parent components above. They are retained in the ordered map form where possible for the sake of easy configuration and usage.

Label Sources

Labels are assigned from multiple sources.

Automatic Labels

Labels which are part of the runtime scaffolding of NoSQLBench are provided for you, as automatic labels. The examples below show standard labels, and an implied nesting of values. Nesting relationships here imply runtime component relationships and lifecycles. For example, you may expect multiple distinct session labels per node label by default, but session uniqueness is only guaranteed within that node label scope.


Notice that some automatic labels are not always included, such as workload, scenario, and step. This is because these labels are only meaningful in the case that you are using named scenarios. It is strongly recommended that in any downstream views you use and require these labels, and encourage your users to use named scenarios as a rule. Ad-hoc activity construction often leads to inconsistency and ambiguity in testing flows, and named scenarios provide a good template format to avoid this.

Metrics, annotations, and logging details can be emitted for any level of labeling, but the labels up to and including session should be considered the minimum set.


Some top-level

Users can inject additional labels in different places:

Label Structure

For an op-specific metrics, the example label set above would be created at runtime as a simple map (shown as JSON for illustration purposes):

  "jobname": "nosqlbench",
  "instance": "default",
  "node": "",
  "session": "nEXHI88",
  "container": "smoketest",
  "workload": "myworkload",
  "scenario": "myscenario",
  "step": "rampup",
  "activity": "rampup",
  "op": "anopname",
  "name": "metricname"

It would be rendered (serialized) in whatever form is idiomatic where it is used.

examples for metric family name and label set in openmetrics




example labels as tags in a grafana annotations API call for an activity:

  "text":"Annotation Description"

Where are label used?


All metrics are labeled according to the label set provided. This supports metric systems which use the Open Metrics exposition format. The recently added push reporter builds this format out of the labels provided where metrics are instanced.

As of 5.21 and newer versions, graphite support is deprecated. This is a necessary change because of the limitations in the graphite metrics naming scheme which directly conflict with modern and more manageable practices for metrics warehousing. Dimensional metrics labels are simply more robust, more expressive, and a more correct way of describing the source of telemetry and logging data.


Annotations are built-in to NoSQLBench, with the most prominent form being the Grafana-specific annotator. When using grafana with another metrics store like Prometheus , or Victoria Metrics , you can have NoSQLBench send metrics to the metrics collector and concurrently send annotations directly to Grafana. This allows populations of metrics data to have markers for each session, scenario, and activity. These annotations embed execution details, timelines, and labeling data so that the metrics from those lifetimes are directly addressable using the self-same label sets.

Grafana annotations have a set of tags which uniquely identify each annotation for lookup. (NOT to be confused with op tags, which allow control of your workload variations.) The labels are provided as values in these tags in conjugate form, such as 'scenario:myscenario234', whereas simple annotation tags would just have a value like 'scenario'. This may cause you to pollute your grafana annotation tag space. For now, we are allowing it to work as-is, to see if time-based filtering and recent views are sufficient to limit the user-facing noise and server-side response.


Logging events sometimes use the labels to identify errors, unique measurements, or configuration details. Even when the label isn't shown as such, there is a direct relationship between a labeled element and it's name. For example, a named scenario "scenario234" would be logged as such, and also annotated with an explicit label set that includes 'scenario=scenario234' via the annotator system.


Labels are inventory addressing systems for runtime scripting. More details on this TBD.

Labeling Standards

In order to make sure that your metrics and annotations have the necessary labels so that the data is traceable, identifiable, and contextual, you may want to set some label validation rules. These rules simply say which label names are required and which are not allowed. If you have a shared system of record for your metrics with other engineers or teams, then some discipline about how these metrics are reported is always required, but not often adhered to, even if there is a shared idea of what the metrics should look like. The net effect is that some of your most useful data gets lost among and indistinguishable from a trove of other metrics data which is extraneous, or worse, your data has this effect on others'.

It is recommended that you have a labeling standard, and that is is somewhere that is easy to share, verify, and enforce. For lack of this, NoSQLBench provides a client-side mechanism to make this possible in practice, although it is opt-in and not enforced from any consuming system. It is expected that NoSQLBench will support loading any provided labeling standards from consuming system automatically in the future.

Some specific examples for why this is important:

Filtering and Validation

example label:


Labeling Specifications

There are two ways to ensure that labels you send are valid:

  1. Adjust the provided labels to conform to the standard before sending.
  2. Validate that the labels provided already conform to the standard before sending.

These two are distinct methods which are complimentary, and thus NB does both. This is because there are label sets which can be aligned to a standard simply by filtering, and there are other cases where this is not possible due to missing data.

Label Specs

These two method are combined in practice into one configuration parameter (for now at least), since the expression of validation and pre-filtering for that validation is effectively one and the same. This is known as the labelspec, and it is simply a one-line pattern which specifies which labels are required, which labels are disallowed, and so on.

Label Filtering

So why would you have a bare label name in lieu of + or vice-versa if they mean the same thing? The reason is that the labelname filter operates as a tri-state filter, going from left to right over the labelspec. You can include, or exclude any label name, and you can also use wildcard patterns simply by using regular expressions. Thus, a spec of lioness,-lion.* indicates that you want to allow lioness labels but disallow any other labels which start with lion. However, this does not say that you require a lioness label, which looks like this +lioness,-lion.*. In both cases, any label which is not mentioned will pass through the filter unchanged. If you wanted to reject all others, you can always end the labelspec with a -.*, such as +lioness, -lion.*,-.*, which is also duplicitous and can simply be reduced to +lioness,-.*.

Label Filtering Effects

The label filter is applied at the point of label usage on outbound signals. Metrics and Annotations both have the filter applied. The underlying label sources are not affected. When a label set is passed through the label filter, some fields may be removed. Otherwise label filtering is entirely passive. The order of labels is preserved if they came from an ordered map.

Label Validation

Label validation occurs on a label set immediately after filtering. The specification is identical, although it weighs a little more heavily in terms of side-effects for validation.

Label Validation Effects

Label validation occurs immediately after label filtering, and will throw an error in your test, with an error that indicates why the label set could not be validated.

This is because invalid labels will cause pain downstream. In most cases, this will manifest early in a test so that you don't have any wasted bake time. It is recommended that you always run a full end-to-end functional validation of all your test workflows before running the more intensive versions. This helps you shake out any issues with late-validation, such as a bad label set in a later stage of testing.

TBD: conditional validation

There is a very clear need to make the validation work a bit more reasonably with regard to labeling context. For example, requiring an activity label at the top level does not work well for the session annotation, nor does requiring and activity label when the scenario label is not present. There are some simple forms which can support these coherently and these will be added shortly.

sketch on conditional validation

  1. appname is a required label name
  2. instance is a required label name and its value must match regex pattern \w+, word chars only
  3. session is a required label
  4. node is a required label
  5. given activity is present as a label name, then k must also be present
  6. given activity is present as a label name, then dataset must also be present
  7. given activity is present as a label name, then dimension must also be present

if a then b a -> b if b then c b -> c if a then b and c a -> b -> c A, thus b! and c!

A! ->b ->c a->b->c +a->b->c

Special cases

Back to top