Observability is often bought as a tool and operated as an afterthought. The architecture question comes first: which signals live together, who runs the pipeline, how long data is kept and what happens when the bill grows faster than traffic. The six models below describe the operating choices available to a small or mid-sized engineering team. They are not ranked, because the right answer depends on skills, risk and the incidents you actually face.
The six models
- Three specialised tools. One product for logs, one for metrics and one for traces, each chosen for depth in its signal.
- Unified commercial platform. A single vendor stores and correlates all three signals, with one query language and one bill.
- Self-managed open source. OpenTelemetry instrumentation feeding open-source backends that your team operates.
- Managed open source. The same open-source APIs and dashboards, hosted and supported by a vendor.
- Cloud-provider native. Logs, metrics and traces delivered by the cloud account that already runs the workload.
- Security-led platform. A SIEM or security analytics platform extended to operational telemetry.
Criteria that decide the comparison
- Correlation: can an engineer move from a trace to the related logs and metrics without copying identifiers between tools?
- Signal depth: does the platform support high-cardinality metrics, structured logs, distributed traces and profiling?
- Cost predictability: is the bill driven by volume, series, hosts, seats or queries, and how does it behave at double volume?
- Operating load: who patches, upgrades, scales and recovers the backend?
- Lock-in: can instrumentation and data move to another backend without a rewrite?
- Residency and access: where does telemetry live, who can read it, and can you meet customer or regulatory expectations?
- Retention flexibility: can you keep recent data hot and older data cheap without losing the ability to investigate?
- Security use: is the same platform expected to support detection and investigation, or only engineering debugging?
- Team skill: does the team have the time and experience to run the model you choose?
Comparison at a glance
| Model | Correlation | Cost predictability | Operating load | Lock-in | Best fit |
|---|---|---|---|---|---|
| Three specialised tools | Weak without shared identifiers | Medium, three contracts | Low to medium | Medium | Teams with deep needs in one signal |
| Unified commercial platform | Strong | Medium, volume and retention sensitive | Low | High if agents are proprietary | Small teams that need one pane of glass |
| Self-managed open source | Strong with OpenTelemetry | High, but labour is real | High | Low | Teams with platform engineering capacity |
| Managed open source | Strong | Medium | Low to medium | Low to medium | Open standards without running the backend |
| Cloud-provider native | Good inside one cloud | Medium | Low | High | Workloads concentrated in one cloud |
| Security-led platform | Good for security events | Low to medium | Medium to high | High | Regulated or threat-heavy environments |
Where each model breaks
Three specialised tools break during an incident. The trace shows a slow dependency, the logs sit in another system with a different time zone and the metric uses a different service name. Engineers spend the first ten minutes translating instead of diagnosing. The model works when a shared identifier, consistent resource attributes and a common clock are enforced across all three.
Unified commercial platforms break on cost and lock-in. A generous retention setting and verbose logs produce a bill that grows without a decision. Proprietary agents make a later migration expensive. They remain the simplest choice for a small team that needs correlation now, provided retention and ingestion are governed from the start.
Self-managed open source breaks when the platform becomes a second product. Upgrades, storage, cardinality and on-call all compete with application work. It is a good choice only when someone owns it explicitly and the team has the skills to recover it at 2am.
Managed open source removes much of that burden while keeping the APIs portable. It breaks when the contract treats retention, cardinality or egress as an afterthought, or when the provider's version lags the upstream project. Ask about upgrade cadence and support boundaries.
Cloud-provider native is convenient and well integrated with the account, identity and network. It becomes expensive or awkward when the business runs across several clouds, when telemetry must outlive a migration, or when the security team needs a single view across environments.
Security-led platforms are built for detection and investigation, not for everyday debugging. Query costs, ingestion models and user experience can be poor for application teams. They are appropriate when security monitoring is the primary use case, not when a SIEM is bought to fix a tracing gap.
Choose by context
- Three tools when one signal is genuinely specialised and the team already shares identifiers across systems.
- Unified platform when a small team needs fast correlation and has no appetite to run a backend.
- Self-managed open source when data control, cost ownership and platform skills all exist in-house.
- Managed open source when you want OpenTelemetry portability without owning storage and upgrades.
- Cloud-native when the workload is concentrated in one cloud and the exit risk is acceptable.
- Security-led when detection and compliance are the driver, with a separate engineering observability path.
Instrument once, route deliberately
Whatever the backend, instrument with OpenTelemetry and use consistent resource attributes: service name, environment, version, region and team. Send telemetry through a collector so routing, sampling, redaction and fan-out are controlled in one place. That layer is what makes a later change of backend a routing decision rather than a rewrite.
Set retention by purpose. Keep enough recent data for incident investigation, move older data to cheap storage and delete what has no legal or operational value. Apply sampling to traces, but keep errors and slow requests. Control metric cardinality before it becomes a budget problem, and never put customer identifiers in labels.
Pilot and score
Run a two-week pilot on three services, including one with a known performance problem. Test a real investigation: start from an alert, move to a trace, find the related log line and confirm the metric that changed. Then test a cardinality spike, a retention query and an export. Ask for a cost estimate at twice current volume.
Score correlation, signal depth, operating load, cost predictability, lock-in, residency and support. Weight them for your team, not for a generic enterprise. A platform that wins on features but requires a new hire to operate is not the cheaper choice.
If you need the collector layer and backend operated for you, start with Managed Technology & Cybersecurity. The pricing page explains how platform and support work are scoped. For the billing side of the same decision, read the observability cost guide.
Sources
- OpenTelemetry, Documentation, accessed 17 November 2026.
- CNCF TAG Observability, Observability Whitepaper, accessed 17 November 2026.
- Google, Site Reliability Engineering: Monitoring Distributed Systems, accessed 17 November 2026.
- Amazon Web Services, Operational Excellence Pillar, accessed 17 November 2026.
Responses (0)
No comments yet. Add the first one below.
Leave a response
Your email address will not be published. Required fields are marked *.