Database Expert · Pro tier

Cassandra

Cassandra is your Database Expert — the guardian of irreversible state. Code can be redeployed; data remembers. She turns migrations, backfills, queries, upgrades and recovery plans into production decisions with a safe path, a credible failure path, measurable abort lines and a proven way home.

Soft voice, hard gates. Cassandra is calm and forensic, never theatrical. Safe work gets SHIP. Sound direction without production evidence gets REHEARSE. A credible path to data loss, corruption, unbounded blocking or an unrecoverable rollback gets BLOCK — followed by the shortest route back to safety.

⛁ Runs the Temporal Twin

Forecasts the safe path, failure path and last cheap exit before a database change reaches production.

⛁ Sets hard gates

Lock, lag, loss and rollback become measurable abort lines — not reassuring prose.

⛁ Scores queries

A deterministic read on index usage, bounding, join and write safety — from the EXPLAIN plan.

⛁ Grounds the verdict

Cloud database inventory, migration code, telemetry, Databricks and deterministic computation when connected.

Cassandra is a Pro-tier specialist. She works on anything you paste — no connection required. Reach her through Sage ("have Cassandra review this schema change") or talk to her directly.

Who Cassandra is#

Cassandra owns stored state: database semantics, schema and query behavior, migrations and backfills, replication, upgrades, backup/restore and recovery readiness. Her recurring frame is Lock · Lag · Loss · Last exit. She labels the evidence she observed, what code computed, what she inferred and what remains unknown. Franz owns data in motion; Cody application code; Kai infrastructure; Daphne security; Finn economics; Paige incident command. Cassandra can consult them without handing the work back to you.

Working with Cassandra#

Hand her the artifact and ask for the verdict. Paste a migration diff, change plan, slow query with its EXPLAIN plan, upgrade proposal or recovery runbook. She asks at most the one question whose answer can change the decision and gives a clearly provisional read while evidence is missing.

Try saying

review this schema-change spec is this migration safe to run in prod tonight? this SELECT is slow — here's the EXPLAIN ANALYZE

Temporal Twin#

For a material DDL, backfill, destructive write or ORM migration, Cassandra runs a deterministic preflight across eight evidence lenses: locking, table rewrites and capacity, old/new application compatibility, data integrity, backfill control, rollback, observability and rehearsal. The result is a 0–100 safety score, an estimated rewrite-time planning floor when size is known, the exact point of no return, and gated EXPAND → MIGRATE → VERIFY → CONTRACT timelines for both the safe future and the future that fails.

Her most valuable sentence is often the last cheap exit. The forecast never connects to or executes against production. Missing engine, version or workload evidence stays visibly provisional rather than being replaced by a confident default.

Try saying

run the Temporal Twin on this 500 GB Postgres migration where is the point of no return in this backfill?

Tech-spec review#

Share a spec involving RDS or database work and Cassandra reviews it for correctness, performance implications, scalability, and operational safety. She calls out the real risks — a missing index that will bite under load, a migration that locks a hot table, a capacity-planning gap, a replication pitfall — and separates blockers from concerns so you know what must change before build starts.

Try saying

review this design's data model for scale will this migration lock the orders table?

Change-management review#

Share a change-management document, runbook, or migration plan and Cassandra evaluates whether it's safe to execute. She backs the read with a deterministic 0–100 safety score across rollback, monitoring, blast radius, window, approvals, clarity, validation and data integrity. When Chronosphere is connected, alarms named in the document are checked against the live monitor catalog; an unavailable catalog remains explicitly unverified, and a monitor that does not exist is treated as a false safety claim.

Try saying

score this change plan — is it safe to run? does this runbook have a real rollback?

Query investigation#

Paste a slow query or query plan and Cassandra runs it through a deterministic performance-and-safety review: it scores sargability and index usage, whether the result set is bounded, join and predicate quality, projection, write/DDL safety, and what the supplied plan actually shows. For read-only SELECTs, a measured plan from representative data can reveal the hot node. For UPDATE, DELETE, INSERT and DDL, she starts with estimate-only EXPLAIN where supported: EXPLAIN ANALYZE can execute the mutation, so measurement belongs on a disposable restored copy or behind an engine-proven rollback boundary.

The classic footgun: a whole-table UPDATE/DELETE with no WHERE holds the verdict at "Risky" — and Cassandra will insist you run the write inside a transaction and check the affected row count before you COMMIT.

Try saying

why is this query slow? [query + EXPLAIN] is this bulk UPDATE safe to run as written?

Version upgrades#

Tell Cassandra your engine and current version — "we're on Postgres 12", "MySQL 5.6", "Mongo 4.4" — and she builds a phased upgrade plan. Her code-backed, versioned catalog maps supported hop patterns, curated breaking-change checks, cutover strategies and downtime estimates into a 0–100 risk score and a runbook with a real rollback plan. The catalog is a preflight, not a substitute for current vendor or managed-service documentation; scheduling decisions are verified against the live support policy.

End of life is a security clock: ask Cassandra for the support status of any version and she'll tell you the GA and end-of-support dates, how far behind the latest you are, and how urgent the move is. She won't schedule a production window until the upgrade has been rehearsed on a staging copy and the rollback rehearsed end to end.

Try saying

plan our upgrade from Postgres 12 to 17, 800 GB, 30-min window what breaks going from MySQL 5.7 to 8.0? is Postgres 12 still supported?

Schema design#

Ask Cassandra about modeling and she gives experience-informed guidance — keys and types, when to normalize and when not to, which indexes actually earn their write cost, and the access patterns a schema should be built around. Concrete, with examples where they help.

Try saying

how should I model a multi-tenant events table? what indexes does this query pattern need?

RDS config & operations#

For the operational side — parameter-group tuning, Multi-AZ vs. read replicas, storage scaling, backup and restore strategy, replication choices — Cassandra names the trade-offs and failure modes. When connected, she can ground the read in database-scoped RDS/Aurora, Azure SQL and Cloud SQL inventory and management-plane change history, then pair it with Grafana or Datadog telemetry and the actual migration code from GitHub.

Try saying

Multi-AZ or a read replica for this workload? what parameter-group changes help this write-heavy load?

Databricks & the lakehouse#

Connect your Databricks workspace in the Connections studio and Cassandra can reach its managed MCP server — Unity Catalog metadata and functions, AI/Vector Search indexes, and Genie natural-language analytics — so her answers are grounded in your actual lakehouse schema, tables and lineage instead of generalities. Unity Catalog enforces its own permissions, so she only ever sees what you've granted, and anything that writes or runs a function goes through Zimac's one-time approval flow. Setup is a paste: your workspace MCP endpoint and a token.

Try saying

what tables are in our sales catalog and how are they keyed? use Genie to pull last quarter's revenue by region

Watches, studios & lessons#

Cassandra shares the team's durable context and navigation toolkit: she can search prior database decisions and recorded evidence pulls, jump into the right studio, and retain a correction such as "always name the last safe rollback point" as a lesson. Recalled context is dated; she never presents it as live state without a fresh connected read.

Try saying

always ask for the plan before scoring a query