In this blog post, we are excited to share the significant strides we have made in improving the UI performance of Elastic APM. Maintaining a snappy user interface can be a challenge when interactively summarizing the massive amounts of data needed to provide an overview of the performance for an entire enterprise-scale service inventory. We want to assure our customers that we have listened, taken action, and made notable architectural changes to elevate the scalability and maturity of our solution.

Architectural enhancements

Our journey began back in the 7.x series where we noticed that doing ad-hoc aggregations on raw transaction data put Elasticsearch^® under a lot of pressure in large-scale environments. Since then, we’ve begun to pre-aggregate the transactions into transaction metrics during ingestion. This has helped to keep the performance of the UI relatively stable. Regardless of how busy the monitored application is and how many transaction events it is creating, we’re just querying pre-aggregated metrics that are stored at a constant rate. We’ve enabled the metrics-powered UI by default in 7.15.

However, when showing an inventory of a large number of services over large time ranges, the number of metric data points that need to be aggregated can still be large enough to cause performance issues. We also create a time series for each distinct set of dimensions. The dimensions include metadata, such as the transaction name and the host name. Our documentation includes a full list of all available dimensions. If there’s a very high number of unique transaction names, which could be a result of improper instrumentation (see docs for more details), this will create a lot of individual time series that will need to be aggregated when requesting a summary of the service’s overall performance. Global labels that are added to the APM Agent configuration are also added as dimensions to these metrics, and therefore they can also impact the number of time series. Refer to the FAQs section below for more details.

Within the 8.7 and 8.8 releases, we’ve addressed these challenges with the following architectural enhancements that aim to reduce the number of documents Elasticsearch needs to search and aggregate on-the-fly, resulting in faster response times:

• Pre-aggregation of transaction metrics into service metrics. Instead of aggregating all distinct time series that are created for each individual transaction name on-the-fly for every user request, we’re already pre-aggregating a summary time series for each service during data ingestion. Depending on how many unique transaction names the services have, this reduces the number of documents Elasticsearch needs to look up and aggregate by a factor of typically 10–100. This is particularly useful for the service inventory and the service overview pages.
• Pre-aggregation of all metrics into different levels of granularity. The APM UI chooses the most appropriate level of granularity, depending on the selected time range. In addition to the metrics that are stored at a 1-minute granularity, we’re also summarizing and storing metrics at a 10-minute and 60-minute granularity level. For example, when looking at a 7-day period, the 60-minute data stream is queried instead of the 1-minute one, resulting in 60x fewer documents for Elasticsearch to examine. This makes sure that all graphs are rendered quickly, even when looking at larger time ranges.
• Safeguards on the number of unique transactions per service for which we are aggregating metrics. Our agents are designed to keep the cardinality of the transaction name low. But in the wild, we’ve seen some services that have a huge amount of unique transaction names. This used to cause performance problems in the UI because APM Server would create many time series that the UI needed to aggregate at query time. In order to protect APM Server from running out of memory when aggregating a large number of time series for each unique transaction name, metrics were published without aggregating when limits for the number of time series were reached. This resulted in a lot of individual metric documents that needed to be aggregated at query time. To address the problem, we've introduced a system where we aggregate metrics in a dedicated overflow bucket for each service when limits are reached. Refer to our documentation for more details.

The exact factor of the document count reduction depends on various conditions. But to get a feeling for a typical scenario, if your services, on average, have 10 instances, no instance-specific global labels, 100 unique transaction names each, and you’re looking at time ranges that can leverage the 60m granularity, you’d see a reduction of documents that Elasticsearch needs to aggregate by a factor of 180,000 (10 instances x 100 transaction names x 60m x 3 because we’re also collapsing the event.outcome dimension). While the response times of Elasticsearch aggregations isn’t exactly scaling linearly with the number of documents, there is a strong correlation.

FAQs

When upgrading to the latest version, will my old data also load faster?

Updating to 8.8 doesn’t immediately make the UI faster. Because the improvements are powered by pre-aggregations that APM Server is doing during ingestion, only new data will benefit from it. For that reason, you should also make sure to update APM Server as well. The UI can still display data that was ingested using an older version of the stack.

If the UI is based on metrics, can I still slice and dice using custom labels?

High cardinality analysis is a big strength of Elastic Observability, and this focus on pre-aggregated metrics does not compromise that in any way.

The UI implements a sophisticated fallback mechanism that uses service metrics, transaction metrics, or raw transaction events, depending on which filters are applied. We’re not creating metrics for each user.id, for example. But you can still filter the data by user.id and the UI will then use raw transaction events. Chances are that you’re looking at a narrow slice of data when filtering by a dimension that is not available on the pre-aggregated metrics, therefore aggregations on the raw data are typically very fast.

Note that all global labels that are added to the APM agent configuration are part of the dimension of the pre-aggregated metrics, with the exception of RUM (see more details in this issue).

Can I use the pre-aggregated metrics in custom dashboards?

Yes! If you use Lens and select the "APM" data view, you can filter on either metricset.name:service_transaction or metricset.name:transaction, depending on the level of detail you need. Transaction latency is captured in transaction.duration.histogram, and successful outcomes and failed outcomes are stored in event.success_count. If you don't need a distribution of values, you can also select the transaction.duration.summary field for your metric aggregations, which should be faster. If you want to calculate the failure rate, here's a Lens formula: 1 - (sum(event.success_count) / count(event.success_count)). Note that the only granularity supported here is 1m.

Do the additional metrics have an impact on the storage?

While we’re storing more metrics than before, and we’re storing all metrics in different levels of granularity, we were able to offset that by enabling synthetic source for all metric data streams. We’ve even increased the default retention for the metrics in the coarse-grained granularity levels, so that the 60m rollup data streams are now stored for 390 days. Please consult our documentation for more information about the different metric data streams.

Are there limits on the amount of time series that APM Server can aggregate?

APM Server performs pre-aggregations in memory, which is fast, but consumes a considerable amount of memory. There are limits in place to protect APM Server from running out of memory, and from 8.7, most of them scale with available memory by default, meaning that allocating more memory to APM Server will allow it to handle more unique pre-aggregation groups like services and transactions. These limits are described in APM Server Data Model docs.

On the APM Server roadmap, we have plans to move to a LSM-based approach where pre-aggregations are performed with the help of disks in order to reduce memory usage. This will enable APM Server to scale better with the input size and cardinality.

A common pitfall when working with pre-aggregations is to add instance-specific global labels to APM agents. This may exhaust the aggregation limits and cause metrics to be aggregated under the overflow bucket instead of the corresponding service. Therefore, make sure to follow the best practice of only adding a limited set of global labels to a particular service.

Validation

To validate the effectiveness of the new architecture, and to ensure that the accuracy of the data is not negatively affected, we prepared a test environment where we generated 35K+ transactions per minute in a timespan of 14 days resulting in approximately 850 million documents.

We’ve tested the queries that power our service inventory, the service overview, and the transaction details using different time ranges (1d, 7d, 14d). Across the board, we’ve seen orders of magnitude improvements. Particularly, queries across larger time ranges that benefit from using the coarse-grained metrics in addition to the pre-aggregated service metrics saw incredible reductions of the response time.

We’ve also validated that there’s no loss in accuracy when using the more coarse-grained metrics for larger time ranges.

Every environment will behave a bit differently, but we’re confident that the impressive improvements in response time will translate well to setups of even bigger scale.

Planned improvements

As mentioned in the FAQs section, the number of time series for transaction metrics can grow quickly, as it is the product of multiple dimensions. For example, given a service that runs on 100 hosts and has 100 transaction names that each have 4 transaction results, APM Server needs to track 40,000 (100 x 100 x 4) different time series for that service. This would even exceed the maximum per-service limit of 32,000 for APM Servers with 64GB of main memory.

As a result, the UI will show an entry for “Remaining Transactions” in the Service overview page. This tracks the transaction metrics for a service once it hits the limit. As a result, you may not see all transaction names of your service. It may also be that all distinct transaction names are listed, but that the transaction metrics for some of the instances of that service are combined in the “Remaining Transactions” category.

We’re currently considering restructuring the dimensions for the metrics to avoid that the combination of the dimensions for transaction name and service instance-specific dimensions (such as the host name) lead to an explosion of time series. Stay tuned for more details.

Conclusion

The architectural improvements we’ve delivered in the past releases provide a step-function in terms of the scalability and responsiveness of our UI. Instead of having to aggregate massive amounts of data on-the-fly as users are navigating through the user interface, we pre-aggregate the results for the most common queries as data is coming in. This ensures we have the answers ready before users have even asked their most frequently asked questions, while still being able to answer ad-hoc questions.

We are excited to continue supporting our community members as they push boundaries on their growth journey, providing them with a powerful and mature platform that can effortlessly handle the demands of the largest workloads. Elastic is committed to its mission to enable everyone to find the answers that matter. From all data. In real time. At scale.

The release and timing of any features or functionality described in this post remain at Elastic's sole discretion. Any features or functionality not currently available may not be delivered on time or at all.

What is Tail Sampling?

Distributed tracing is useful for debugging, but at production scale it comes with processing overhead and storage costs, at which point sampling becomes a natural way to maintain the value of tracing while keeping costs under control. Tail-based sampling, or tail sampling, is a technique that makes a sampling decision conditionally at a later stage, so that high-value traces like errors or slow transactions are more likely to be sampled. The opposite is head sampling, which makes the decision at the start of a trace, before any such information is available.

How does the tail sampling processor work?

The tail sampling processor buffers 100% of incoming traces (or spans, used interchangeably), then forwards the sampled subset after applying the sampling policies. Buffering is a major source of memory usage, and it scales proportionally to the volume of spans, a well known pain point in the community.

Memory usage is bounded by configuration parameters like decision_wait and num_traces. Setting decision_wait to 1 minute means a sampling decision is made for a trace after 1 minute, during which all spans for that trace are expected to have arrived. If a trace is slower than 1 minute, the decision is made with some spans missing.

As a side note, scaling out the tail sampling setup involves using the loadbalancingexporter to satisfy the requirement that all spans for a trace must be routed to the same collector. This introduces some operational complexity and potentially data loss during collector restarts. But this post focuses on the memory usage of a single tail sampling processor instance, regardless of horizontal scaling.

Why does tail sampling cause memory pressure?

These parameters introduce a tradeoff between data loss and memory usage, and they require assumptions about the shape of traces: how slow they can be, how many spans they contain, how large each span is. These assumptions can become stale as instrumentation evolves.

How much data loss is acceptable to limit memory usage, and can the tradeoff be improved? The following two contributions aim to give operators more flexibility.

How span-ingest reduces tail sampling memory by releasing spans early

sampling_strategy is a new configuration option added to the tail sampling processor in v0.149.0.

sampling_strategy defaults to trace-complete, which matches the original behavior: sampling policies are only evaluated when decision_wait has elapsed, at which point the trace is considered complete. (There is a similar config, decision_wait_after_root_received, for optimization, but it is excluded from this discussion for simplicity.) This means all spans are buffered in memory for roughly decision_wait before being released, regardless of whether a decision could have been made earlier. For example, health check spans that should always be dropped are still held in memory until policy evaluation time.

Alternatively, sampling_strategy can be set to span-ingest, where spans are evaluated individually at ingest time. This allows terminal decisions, specifically drop or sampled, to be made earlier, freeing memory by dropping or exporting all spans buffered so far for that trace before decision_wait elapses. In the health check example, a policy can be configured to drop the entire trace as soon as the root span belongs to a health check. It is worth noting that an unsampled decision, unlike an explicit drop, is not terminal, as it can be overruled by a sampled or drop decision from another span in the same trace, so unsampled traces cannot be released early.

Switching from trace-complete to span-ingest will require policy adjustments, as policies can no longer assume all spans are available at evaluation time. Moreover, not all policy types are supported with the span-ingest strategy.

Disk-backed tail sampling storage with Pebble

Even with span-ingest, all spans are still buffered in memory. As decision_wait is increased to accommodate slow traces and num_traces is increased to limit data loss, the collector will eventually hit its memory limit and get OOM killed, resulting in further data loss.

What if traces were buffered on disk instead, where there is an order of magnitude more capacity? The main drawback is performance: disk throughput and latency, even with SSDs, are at least an order of magnitude slower than memory, so disk writes need to be efficient. For this reason, Pebble, an LSM database, was chosen as the storage backend for its fast write performance. Read performance is less of a concern, as reads only happen for the sampled subset of traces when sampling_strategy is set to span-ingest.

The implementation introduces a TailStorage interface for trace storage operations, and a new tail_storage option in v0.150.0 (behind feature gate processor.tailsamplingprocessor.tailstorageextension) to configure the storage backend. The default in-memory behavior is unchanged, but it is now possible to swap in a different storage backend, like the new pebbletailstorageextension contributed to Collector Contrib.

Tail sampling memory benchmarks: trace-complete vs span-ingest with Pebble

Benchmark setup: OpenTelemetry Demo with fan-out collectors

The following benchmarks were produced by running OpenTelemetry Demo with increased load against a pipe collector, which receives all spans and fans them out to two identical collectors under observation (CUO-A and CUO-B), differing only in their tail sampling configuration. Measurements include pipe collector throughput, spans received, spans sent (sampled), CPU usage, and memory usage.

Benchmark setup diagram

                      demo ns
     +----------------------------------------+
     |  opentelemetry-demo                     |
     |    loadgenerator (locust)               |
     |    services: frontend, cart, ...        |
     |    demo-collector                       |
     +----------------------------------------+
                          |  OTLP/gRPC
                          v
                     chamber ns
     +----------------------------------------+
     |             pipe-collector             |
     |         receive once, fan out          |
     |      exporters: [otlp/a, otlp/b]       |
     +----------------------------------------+
              | OTLP                  | OTLP
              v                       v
     +----------------+      +----------------+
     |     CUO-A      |      |     CUO-B      |
     | tail_sampling  |      | tail_sampling  |
     |   (config A)   |      |   (config B)   |
     +----------------+      +----------------+

Tail sampling processor configurations

CUO-A

config:
  processors:
    tail_sampling:
      sampling_strategy: trace-complete
      decision_wait: 5m
      num_traces: 5000000
      block_on_overflow: true
      decision_cache:
        sampled_cache_size: 10000
        non_sampled_cache_size: 200000
      policies:
        - name: root_1pct
          type: and
          and:
            and_sub_policy:
              - name: root_span_only
                type: ottl_condition
                ottl_condition:
                  error_mode: ignore
                  span:
                    - "IsRootSpan()"
              - name: root_probabilistic
                type: probabilistic
                probabilistic:
                  sampling_percentage: 1.0

CUO-B

CUO-B uses the same tail sampling processor configuration as CUO-A, except it sets sampling_strategy: span-ingest and tail_storage: pebble_tail_storage/main, along with its corresponding pebbletailstorageextension configuration.

extensions:
  pebble_tail_storage/main:
    directory: /var/lib/otelcol/pebble

Memory, CPU and throughput results

The following tables compare trace-complete (CUO-A) against span-ingest with Pebble disk storage (CUO-B) across memory, CPU and throughput. The process RSS, Go heap allocation, and per-process CPU measurements come from OpenTelemetry Collector internal process and runtime metrics, while container working set and container CPU come from Kubernetes cgroup metrics scraped by kubelet/cAdvisor.

Memory (peak over the window)

CPU (total over the window)

Throughput (total over the window)

Tail sampling

• cuo-a = trace-complete, cuo-b = span-ingest-pebble
• Window: 15m 39s (t+0:00 start, t+10:06 drain start, t+15:39 drain end)

The results show a significant memory reduction when using pebbletailstorageextension with span-ingest, at the cost of increased CPU usage from event serialization and database overhead.

What's next for OpenTelemetry tail sampling

Both sampling_strategy and pebbletailstorageextension are still in their early stages at the time of writing. Feedback and contributions are welcome in the OpenTelemetry Collector Contrib repo. Stay tuned for more improvements.