Database monitoring: what it costs, and what it costs your database
This page states two things. What you are charged for monitoring a Postgres database, and what monitoring charges your database. The second is a number Bucker measures on your database and shows you, not a benchmark taken once somewhere else.
Status. The statement of what is and is not metered is in force. The weights of the things that are weighted, and the free allowance, are not priced yet: they wait on the owner's decision OD-DP-06 (weights, not meters, with a safety control on every plan). Where a number is missing below, it is missing on purpose, and this page will not guess one.
What you are charged for
There is no per-database-host fee, no per-replica fee, no per-table fee and no per-seat fee. A replica, a branch of a serverless database or a second database in the same project adds no unit to your bill by existing. There are no credits that expire.
| What Bucker does | Charged |
|---|---|
| Statement rollups (the per-fingerprint, per-hour statistics) | No. Unmetered, as uptime is. |
| Advisors and the findings they raise | No. |
| The coverage matrix (what each signal can and cannot see, and why) | No. |
| Queue and cache health, and every gauge Bucker pulls from your database | No. They are derived rows written once a minute. |
| Raw evidence samples (activity samples, plan evidence) | Yes, on the weighted unit budget at the log weight, past a free daily allowance per database. The allowance and the weight are unpriced until OD-DP-06. |
| A kept, validated index change | May be an outcome in the verified-fix lane. Whether and at what price is the owner's decision (OD-DP-06); nothing is charged for it today. |
| An explain an agent runs against your database | Its cost to your database is charged to CUSTOMER_DB_TIME (the budget that limits what Bucker may spend of your database's time), not to your bill. |
Going over a quota degrades with a receipt and is never silently billed. No plan charges for the unmetered rows above, and no agent can change a plan or a price: placement on a plan and the outcome price are owner decisions.
What monitoring costs your database
Bucker's collector is built not to become your outage. It reads over one connection, in a read-only transaction, under a 2 second statement_timeout and a 100 ms lock_timeout, with application_name set to bucker-collector.
The published budget, per database
| Limit | Value |
|---|---|
| Connections held | 1 |
| Statements per minute (busiest minute of the last 5) | 30 at most |
| Time your database spends on the collector's statements per minute | 3 seconds at most (5 % of one connection) |
statement_timeout on every collector statement |
2 seconds |
| Target share of your database's own time | 0.5 % or less |
| Pass | Cadence |
|---|---|
| Statement statistics | every 60 s; statement text fetched only for query ids not seen before, at most 200 per pull |
| Activity | every 10 s |
| Blocking | every 15 s, only while waits exist |
| Explains | at most 10 per fingerprint and 60 per database per hour |
| Schema and settings | every 6 hours |
| Column statistics | every 24 hours |
The same numbers are returned by db_connection_get as collectorLoad.budget, so an agent can read them.
The cap, and what the collector does above it
Every session and every statement the collector sends is counted per database. When the busiest recent minute is over the statements cap, the time cap or the one-connection limit, the collector backs off on that database:
- it opens no session until the wait is over (2 minutes after a first breach, doubling on each consecutive breach, to at most 30 minutes);
- it reports it:
db_connection_getshowscollectorLoad.currentwithstate: "backing_off", which limits were exceeded, the time it will ask again and how many breaches in a row; - it does not treat it as an error, does not open the connection's breaker and does not delete anything;
- a window that is clean afterwards ends it.
One database backing off never slows another. The meter lives in the worker process that collects, so a restart forgets a breach and the collector measures again from scratch.
The measured cost
collectorLoad.measured is read from the pg_stat_statements rows of the collector's own role: the database time that role used, per hour, and its share of the database's total. It is shown beside the budget.
- It is an upper bound. The setup script creates a role the collector alone uses; if you also point an application at that role, its time is counted as Bucker's.
- Hours with nothing recorded say so (
null), never0 %. - It reads the hour rows Bucker keeps, which are kept for 14 days.
withinTargetcompares the share over the whole read with 0.5 %.
The product's own goal is that the collector's database time is at or below 0.5 % of total on 95 % of connected databases, published per database. Datadog's published figure for its Agent on an RDS m5.large is about 1 % of database CPU time, benchmarked once; Bucker's is measured on yours.
What is decided and what is not
| Question | State |
|---|---|
| No per-host, per-replica, per-table or per-seat fee | Decided. |
| Rollups, advisors, coverage, queue and cache health and pulled gauges unmetered | Decided. |
| Weight of raw evidence samples, and the size of the free daily allowance | Unpriced, OD-DP-06. |
| Whether a kept index change is an outcome with a price | Unpriced, OD-DP-06. |
| The collector's cap numbers above | Published and enforced now. |