Skip to main content
Version: Next

SigNoz resource profiles

These SigNoz sizing overlays provide example configurations to size the bundled SigNoz chart. Use them as starting allocations for workload assessment.

ProfileClickHouse layoutSigNoz collectorsMetadata storeIntended use
Small1 shard, 1 replica1SQLite by default; PostgreSQL optionalEvaluation
Medium1 shard, 1 replica1SQLite by default; PostgreSQL optionalCustom deployment with larger vertical allocations
Distributed2 shards, 1 replica per shard3SQLite by default; PostgreSQL optionalAdaptation of the upstream Community resource reference

The distributed profile follows the component counts and CPU/memory allocations in SigNoz Community resource planning. The example sets both requests and limits to the reference allocation for those components. Small and medium provide lower resource starting points for OBaaS; validate their suitability against your workload.

Distributed availability and routing​

The distributed example spreads data over two ClickHouse shards with one replica per shard. It provides horizontal capacity, no ClickHouse replica redundancy. Loss of either shard can interrupt ingestion and complete queries; permanent volume loss can lose that shard's data. Three ZooKeeper instances provide coordination quorum, not additional copies of ClickHouse data. The SigNoz server and its default SQLite metadata store also remain single instances. This profile is not a highly available deployment.

SigNoz distributed ClickHouse guidance illustrates two replicas per shard plus three ZooKeeper instances. Treat that as a separate availability design: budget resources and storage per replica, place replicas across failure domains, verify backup/restore and failover, and plan migration before changing an existing layout. That upstream page uses an older schemaMigrator key; this bundled chart uses telemetryStoreMigrator.

The example moves Prometheus scraping and Kubernetes events to the singleton k8s-infra.otelDeployment. Its outage creates a collection gap even if all three central collectors are healthy. Do not increase this singleton's replica count without assigning disjoint scrape targets and event collection ownership; otherwise telemetry can be duplicated. Central collector replication does not replicate its work.

Three central collector replicas also do not guarantee balanced ingestion. A Kubernetes Service balances connections, and long-lived OTLP/gRPC connections can remain pinned to one pod. Measure accepted telemetry, queues, memory and CPU per pod. Consider client-side gRPC balancing or a gRPC-aware proxy only after testing the client, discovery, connection churn and failure behavior. See OpenTelemetry scaling guidance.

The pinned SigNoz collector v0.144.10 span-metrics implementation aggregates individual spans in local maps, resets aggregates for delta export, and adds signoz.collector.id to dimensions. The bundled configuration selects AGGREGATION_TEMPORALITY_DELTA and retains that collector dimension. This source check supports keeping the current settings; it is not an end-to-end validation of multi-replica queries. Do not remove collector identity, switch temporality, or prescribe trace-ID affinity from generic stateful-processor advice alone. Before changing routing, distribute known spans across replicas and compare service/operation call counts, errors and merged histogram counts/buckets with a single-collector baseline, aggregating across collector IDs. Exercise reconnects and pod replacement. Verify metrics routing separately, especially cumulative streams, and avoid replaying the same spans to multiple collectors.

Resource allocations​

Each cell shows CPU cores / memory GiB, as request → limit, per instance. Memory expressed in MiB in YAML is shown as fractional GiB here.

ComponentSmall (Requests → Limits)Medium (Requests → Limits)Distributed (Requests → Limits)
ClickHouse, per shard2 / 8 → 4 / 164 / 16 → 8 / 3216 / 32 → 16 / 32
SigNoz collector, each0.25 / 0.5 → 1 / 21 / 2 → 2 / 44 / 16 → 4 / 16
SigNoz server0.25 / 0.5 → 1 / 10.5 / 1 → 2 / 24 / 8 → 4 / 8
ZooKeeper, each0.25 / 2 → 1 / 40.25 / 2 → 1 / 42 / 8 → 2 / 8
Node agent, per node0.1 / 0.25 → 0.5 / 0.50.25 / 0.5 → 1 / 10.25 / 0.5 → 1 / 1
Cluster collector, singleton0.25 / 0.5 → 1 / 10.5 / 1 → 2 / 21 / 2 → 2 / 4

Totals​

These totals include all ClickHouse, SigNoz collector, SigNoz server and ZooKeeper instances, the ClickHouse operator and its metrics exporter, and the singleton cluster collector. Each cell shows requests → limits.

ProfileCPU coresMemory (GiB)
Small3.2 → 911.875 → 24.75
Medium6.45 → 1622.375 → 44.75
Distributed55.2 → 57146.375 → 148.75

Configuring your cluster​

Run from the repository root after preparing OBaaS prerequisites, database, access, secrets and cluster capacity. Replace the release, namespace and values placeholders with the settings for your installation.

helm lint helm/infra-charts/obaas \
-f <environment-values.yaml> \
-f helm/infra-charts/obaas/examples/values-signoz-small.yaml \
-f <storage-and-site-overrides.yaml>

helm template <app-release> helm/infra-charts/obaas -n <application-namespace> \
-f <environment-values.yaml> \
-f helm/infra-charts/obaas/examples/values-signoz-small.yaml \
-f <storage-and-site-overrides.yaml> > /tmp/obaas-rendered.yaml

helm install <app-release> helm/infra-charts/obaas \
-n <application-namespace> --create-namespace \
-f <environment-values.yaml> \
-f helm/infra-charts/obaas/examples/values-signoz-small.yaml \
-f <storage-and-site-overrides.yaml>

Substitute medium or distributed as appropriate. Inspect rendered files securely: Later files override earlier ones, and YAML lists replace entire lists. Merge custom receivers, processors and RBAC deliberately.

Storage, retention and buffering​

Illustrative ClickHouse PVC capacities are 200Gi (small), 500Gi (medium) and 1Ti per shard (distributed). ZooKeeper uses 20Gi per instance. SQLite, when used, has 10Gi. Select storage classes, volume performance and capacity in an overlay. For example:

signoz:
clickhouse:
persistence:
storageClass: <clickhouse-storage-class>
size: <measured-capacity-per-instance>
zookeeper:
persistence:
storageClass: <zookeeper-storage-class>
signoz:
persistence:
storageClass: <metadata-storage-class>

Use measured compressed storage growth across telemetry and dependent tables, desired retention, shard distribution and replication to estimate capacity. Allow additional room for merges, diagnostics, delayed TTL cleanup and recovery. Separate PVCs on the same exhausted backing filesystem do not isolate failures. Shrinking PVCs is not a supported way to move down a size profile.

The included diagnostic TTL settings do not configure application retention. Choose application logs, traces and metrics retention independently (for example, 2, 7 or 15 days) using the version-appropriate SigNoz retention controls. Verify dependent-table TTLs, existing-data behavior and disk reclamation before claiming that a retention policy bounds storage.

All three collector tiers set GOMEMLIMIT and a memory_limiter, with headroom below their container limits. The central collector also bounds batch sizes. Starting values for the additional tiers are:

Tier / profileContainer limit (MiB)GOMEMLIMITLimiter hard / spike (MiB)
Node / small512300MiB384 / 64
Node / medium or distributed1024600MiB768 / 128
Cluster / small1024600MiB768 / 128
Cluster / medium20481200MiB1536 / 256
Cluster / distributed40962400MiB3072 / 512

The soft refusal threshold is hard limit minus spike allowance. The Go runtime limit is a soft target, not a cap on total process RSS. Tune all three settings together under load.

Memory pressure can refuse or lose telemetry at the agents and singleton before it reaches a central persistent queue. Pull receivers and file readers have different retry/checkpoint behavior; memory limiting cannot guarantee delivery. SDK/agent retries can expire.

Exporter queues and retries retain chart defaults. These examples do not add persistent queue storage or promise restart-surviving buffering. For small or medium, the optional persistent queue overlay provides a dedicated PVC for a single collector.

Existing releases and validation​

These examples describe fresh installations. Before changing an existing release, preserve its effective values, inventory versions and topology, back up telemetry and metadata, and test restoration. Establish data placement, distributed query coverage, schema compatibility, downtime and recovery before changing shards or replicas. Enabling PostgreSQL does not copy SQLite metadata. A Helm rollback alone does not reverse database migrations. Review the OBaaS upgrade guidance and the upstream procedures for the installed versions.

Helm linting and rendering validate chart compatibility only. Before using a profile operationally, measure sustained/burst ingestion with representative payloads, cardinality and concurrent queries; reconcile generated and queryable telemetry; measure freshness, queue drainage, memory peaks, throttling, query latency and disk growth. Exercise backend slowdown, collector restart, disk pressure and recovery. Use those measurements to choose resources, storage and operational targets for your environment. Kafka is separate and is not enabled by these overlays.