Database scope: what Bucker will not do to your databases, queues and caches
Draft, awaiting the owner's ratification of version 1.0. The decisions behind it (taken 2026-10-28) are listed at the end; the ratification is recorded against a version of this page, and none is recorded yet.
Bucker connects to no customer database, queue or cache today. This page is published first, before
anything in the area ships, so that the boundary is on record and an agent or a buyer expecting an
AI database administrator finds out in one read that there is none. It is version 1.0, as of
2026-10-28, and its table is also kept as not-offered.json in the database module so that a read
can serve the same list: an agent asked to resize a database answers "Bucker does not resize
databases; here is the saturation fact and the provider's console" instead of reaching for a generic
tool.
What Bucker does in this area
Bucker observes databases, queues and caches: it reads facts (usage beside a limit, a query's fingerprint joined to the traces that issued it, queue age and depth, hit rates), prints the threshold behind every finding, and turns a finding into an ordinary issue. It acts on a customer's data store in three narrow ways, each requested by an agent or a person and decided by a person:
- Cancel one statement (
pg_cancel_backendon one identity-pinned session). Never terminate. - Redrive dead-letter messages, paced and capped. Purging is a human act under dual control.
- Propose an index or a timeout as a pull request or a copyable statement. Applying an index itself comes later, under dual control and with an automatic revert; applying a parameter never comes.
All of this is built in stages and none of it is built yet. Everything on this page is the other side of that line.
What is not offered
Each item has no action name, so no policy row can enable it. The kept out by column is the line or decision that keeps it out; they are defined at the end of the page.
| Not offered | Area | Why | Kept out by | Instead |
|---|---|---|---|---|
| Capacity planning | Database | Sizing advice for a database, queue or cache is the infrastructure AI-SRE the line bars. Bucker reports usage beside a limit, with the threshold printed, and a projected breach time where a rule reads the series; it never says how large an instance should be. | DNB 2, OD-DP-02 | The saturation fact beside the error spike, and the provider's own console for sizing. |
| Resizing an instance or changing its plan | Database | A resize is a capacity action on infrastructure, and most providers cannot undo one cleanly. No lever Bucker offers acts on nodes, networks or capacity. | DNB 2, OD-DP-01, OD-DP-21 | The saturation fact and the provider's console; a person resizes. |
| Failing over to a replica | Database | Failover moves production traffic and risks split-brain; it is an infrastructure action Bucker never takes or requests. | DNB 2, OD-DP-02, OD-DP-21 | Replication lag and connection facts as evidence, and the provider's failover control for a person. |
| Creating, promoting or deleting replicas, backups or restore points | Database | Replica and backup management is infrastructure operation. Bucker holds a read-only view of a database and no credential that could do any of it. | DNB 2, OD-DP-02 | Whether a replica is connected is reported as coverage; the provider's console manages them. |
| Changing database parameters, with or without a model deciding the value | Database | Parameters are never applied by Bucker. A change to one is human-only and critical and goes out as a pull request or a statement a person runs. The name db_parameter.apply is reserved, human-only, and nothing performs it. |
OD-DP-21, DNB 2 | A proposed timeout as a draft pull request or a copyable statement, with the replayed counts that justify it. |
| Running VACUUM or changing autovacuum settings | Database | A vacuum or an autovacuum setting is a parameter or a maintenance command on a customer's database. Bucker reports the lag and its cause; a person runs the command. | OD-DP-21, OD-DP-02 | The vacuum-lag finding with its thresholds; a write pattern that causes bloat goes to the code-fix lane. |
| Terminating a database session (pg_terminate_backend) | Database | Terminating a backend drops a connection and rolls back its transaction, and cannot be undone. Never offered, to anyone. Only cancelling the current statement of one identity-pinned session is offered. | OD-DP-21 | Cancel (db_session.cancel, pg_cancel_backend only, requested by an agent and approved by a person), or a timeout proposal for the role. |
| Forcing a query plan (hints, plan pinning, planner settings) | Database | Forcing a plan changes how a customer's database executes every matching statement and is a parameter-class change. Bucker shows the plan and the evidence; it does not pin one. | OD-DP-21 | Planner-only evidence (EXPLAIN without ANALYZE, hypothetical indexes) and an index or rewrite as a pull request. |
| Kafka administration: creating or deleting topics, fetching messages | Queue | Topic changes are infrastructure administration with no inverse, and fetching a message reads end-user payload. Bucker reads queue depth and age, not the broker's control plane or its messages. | DNB 2, OD-DP-02 | Queue depth, age and consumer-lag facts from the queue's own metrics. |
| Data Streams pathway checkpoints (a tracer in every producer and consumer) | Queue | Pathway checkpoints need a Bucker-owned library in every service's message path and a proprietary context header, which is a traffic recorder of Bucker's own. Open standards are used instead. | DNB 4, OD-DP-03 | W3C trace context and span links, so per-hop time in a queue comes from the spans a service already sends. |
| Tracing Spark, Databricks, EMR, Airflow, dbt or Glue jobs | Pipeline | Job tracing instruments a customer's compute platform from inside it, which is infrastructure monitoring of a kind Bucker's customers do not run through Bucker. Job liveness is a cron monitor. | DNB 2, OD-DP-02 | A cron or heartbeat monitor on the job, and its alerts routed to the team that owns it. |
| Lineage and a data catalog over a customer's warehouse | Warehouse | Bucker connects to no customer warehouse (the Snowflake and BigQuery drivers wait for a buyer) and offers no SQL surface in the first version, so it holds no lineage or catalog of one. The lineage of Bucker's own exports belongs to the warehouse area, not to databases. | OD-DP-24, OD-DP-04 | The warehouse vendor's own catalog and lineage. |
| Flushing a cache (FLUSHALL, FLUSHDB) | Cache | A flush destroys every key, has no inverse and can turn a hit rate into an outage. It is never offered, and no policy row can enable it because it has no action name. | DNB 2, OD-DP-02, OD-DP-21 | The hit-rate drop, the endpoints that missed and the release whose deploy window precedes it, as one issue. |
| Deleting cache keys | Cache | Deleting a key is an unrecoverable write to a customer's data store. Bucker reads no cache keys or values; a key name is kept only if the SDK sends it, as data. | OD-DP-21, DNB 2 | The lookups that missed and the fingerprint that issued them; a person deletes keys. |
| Changing a cache's eviction policy or memory limit | Cache | Eviction policy and memory limits are parameters, and a memory change is a capacity action. Neither is applied by Bucker. | OD-DP-21, DNB 2 | Eviction and memory facts beside the hit-rate drop, with the provider's console for the change. |
| Pausing a database or project at its provider (for example a Supabase pause) | Provider | A pause takes a customer's database offline and some providers offer no way back. It is an availability and capacity action, not one of the three bounded workload levers. | DNB 2, OD-DP-01, OD-DP-02 | Whether the database is paused is reported as coverage; the provider's console resumes it. |
| Restoring a database or a branch (point-in-time restore, Neon branch restore) | Provider | A restore overwrites current data and cannot be undone. It is infrastructure operation with no inverse. | DNB 2, OD-DP-01, OD-DP-02 | The migration and schema-change evidence that explains why a restore is wanted; a person restores. |
| A chat box that investigates a database | Agent | No conversational copilot is built anywhere: a mention may compile a closed phrase and answer, but may not plan over tools or assert a cause an RCA did not produce. | DNB 6, OD-DP-04 | Findings with their thresholds as ordinary issues, and typed reads an agent can call. |
| A Bucker-run synthetic Postgres that benchmarks a rewrite | Database | Building a copy of a customer's database from its schema and statistics and running it is a sandbox of Bucker's own, which the line bars. Planner evidence is gathered inside the customer's database instead. | DNB 4, OD-DP-03 | EXPLAIN without ANALYZE and hypothetical indexes run inside the customer's database, as evidence for an index or rewrite pull request. |
| Capturing raw query parameters or literals | Database | Literals and bind parameters are end-user values. A statement is stored normalized and never with its parameters, and a permission to capture them does not make row values safe to hold; widening collection is a privacy-posture decision, not a toggle. | OD-DP-20 | The normalized statement, its fingerprint and the plan without constants. |
| Paging a person or escalating on a queue alert | Queue | Schedules, escalation policies and internal incident objects are an on-call product, which is not built. A queue alert is routed through the channels the organization already has. | DNB 7, OD-DP-05 | An alert rule on queue age or depth, delivered to the organization's own paging tool. |
How the list is kept honest
- Every item cites a decision. A test requires at least one citation on each, requires each to be a real line or decision, and requires that every line and decision cited here is defined below.
- Nothing can perform an item. The same test reads the action registry and the tool list and
fails if an action or a tool in this area could resize, fail over, terminate, flush, restore,
pause, tune, force or reboot anything. The one name that comes near,
db_parameter.apply, is reserved, human-only and performed by nothing. - The page and the list are one file. The table above is generated from
not-offered.json, and a test fails when they differ.
The decisions cited
Bucker builds against a short list of things it will not build (the do-not-build lines, DNB) and records the owner's decisions on them (OD-DP). These are the ones this page rests on.
- DNB 2. No general-purpose infrastructure AI-SRE: no diagnosing or acting on Kubernetes, network or capacity.
- DNB 4. No sandbox, eBPF profiler or traffic recorder of Bucker's own; open-source tools only, through a seam.
- DNB 6. No conversational copilot.
- DNB 7. No on-call or escalation product, and no SLA.
- OD-DP-01. Three bounded workload levers (restart, scale within a band a person set, roll back one revision), none on nodes, networks or capacity.
- OD-DP-02. DNB 2 bars diagnosing and acting on infrastructure, not observing it; the published out-of-scope statements are the line's public text.
- OD-DP-03. Where DNB 4 bends: open-source tools through a seam, and nothing of Bucker's own.
- OD-DP-04. How far DNB 6 reaches: closed phrases per surface, and no SQL text in the first version.
- OD-DP-05. DNB 7 allows six narrow slices that are not on-call.
- OD-DP-20. The privacy posture: no default widens what is collected.
- OD-DP-21. Database write actions: cancel first, an applied index later, parameters never; terminating a session, failing over and resizing are never offered.
- OD-DP-24. The deferred connector line: the Snowflake and BigQuery drivers wait for a buyer.
The network equivalent of this page is infrastructure scope. The trust page links both.