Security and observability data

Most logs are evidence, not alarms. Stop paying alarm prices for evidence.

Keep everything in a lakehouse you own, and query it where it lives.

Kevin Keller · kevinkeller.org

The problem

The bill grows with the business, whether anyone reads the data or not

  • SIEM and observability tools are often priced by the data they take in
  • More services, more cloud accounts, more AI agents: more logs
  • The usual fix, cutting sources or retention, quietly deletes the evidence

Three forces

Everything pushes the same way

  • Volume grows: pricing by volume turns growth into cost
  • Most data is rarely read: DNS, flows, audit trails, debug logs
  • Retention goes up: 30 to 90 days in the SIEM, one to seven years asked by compliance
  • The result: budget versus evidence, a false choice

The idea

One record, two speeds.

Separate the system of record from the systems of reaction.

Two tiers

Pay for the tier each kind of data needs

Real-time tools (SIEM, APM)Open lakehouse
What goes inThe time-critical sliceEverything
ForAlerts within minutesInvestigations, compliance, AI
RetentionDays to weeksAs long as policy requires
Query speedMillisecondsSeconds

The building blocks

Four parts, none tied to one vendor

  1. A routing layer that decides where each event goes.
  2. Open storage you own, such as Apache Iceberg on object storage.
  3. Federated queries from the tools analysts already use.
  4. Detections on the lake where seconds are enough.

Deciding what goes where

Does this event need someone to react within minutes?
  • Yes: endpoint alerts, suspicious logins, privilege escalation, on-call errors
  • No: web access logs, DNS, network flows, cloud audit trails, debug logs
  • Start conservative and reduce as confidence grows

What it's worth

70 to 80%

lower monthly cost in one worked example

100 GB of security logs a day: roughly $15,000 to $20,000 a month with everything in the SIEM, roughly $3,500 to $6,500 with the critical slice there and the rest in the lake.

One scenario's estimates, not a quote

Beyond the licence line

What leaders get

  • Retention you can afford: five years became practical
  • Better investigations: months of history, not 30 days
  • One copy, many uses: security, operations, finance, AI
  • Negotiating power: your history isn't locked in a tool

A phased plan

Nothing is taken away before the replacement works

  1. Measure ingest by source and by how it's used.
  2. Build the lake and routing, alongside today's setup.
  3. Connect federated search for analysts.
  4. Move high-volume, low-urgency sources one at a time.
  5. Renegotiate the licence for the real-time slice.

When it doesn't pay off

Be honest about the fit

  • Strongest above about 50 GB a day, not justified below about 10
  • Detections that need every source in one place: audit them first
  • Two platforms to run, and someone must own the routing rules
  • The lake holds sensitive data and needs the same controls

The takeaway

Make ingest volume an architecture decision, not a licence surprise.

Read the full article · The executive brief

Kevin Keller · Figures are estimates for one scenario · Not an official Snowflake solution