Cost & Performance
Find and cut warehouse spend
Consumption rates V2 | Consumption rates V1
Bayes is Monte Carlo’s AI teammate. It observes your agents and the data that powers them, finds what breaks, proposes and verifies fixes, and gets better at your environment over time.
Bayes finds storage and compute waste in your warehouse — storage (tables you're paying to keep but no longer use) and compute (expensive or inefficient query patterns) — so data teams can reduce spend. You can explore with Bayes in conversation, browse a prioritized Findings report of your top optimization opportunities, or query Bayes through the Agent Toolkit (MCP).
What makes it different from a standalone cost tool: Monte Carlo already maps your lineage, runtime, and query history. So Bayes doesn't just tell you what's expensive — it tells you what's safe to change, because it can see what depends on each table before suggesting you touch it.
When Bayes analyzes cost and performance, it proposes changes; it does not act on your behalf. It surfaces the finding, the reasoning, and the change to make — you stay in control of what actually changes in your environment.
How it works
Bayes analyzes the metadata Monte Carlo already collects — table sizes, read/write history, lineage, and query logs — and turns it into a ranked list of key findings: specific, actionable opportunities to reduce spend. Each finding names the assets involved, explains why it's flagged, estimates the impact where it can, and suggests what to do.
You can also click "Find out more" on any finding to open it in a conversation with Bayes, with full context preloaded — from there, ask any follow-up question to investigate deeper.
Because it reasons over lineage, Bayes can distinguish a table that's genuinely orphaned from one that quietly feeds a downstream view or dashboard — the difference between a safe cleanup and a broken pipeline.
Capabilities
Alongside the Queries tab, where you can rank every query by run time or credits and dig into each one yourself, Bayes covers three areas of cost analysis — surfacing what's worth acting on without you having to go looking.

Storage Cost Analysis
Bayes identifies tables wasting storage spend by classifying them into waste patterns, including:
- Unread tables — exist but are never queried by downstream consumers
- Write-only tables — continuously written to but never read
- Dead-end tables — no downstream lineage (nothing consumes them)
- Static waste — data hasn't changed in a significant period
- Zombie tables — no longer written to, but still consuming storage
- Other stale — other signals of disuse
For each candidate, Bayes uses Monte Carlo lineage to verify the table is truly unused before proposing cleanup, and estimates the storage savings from removing it. Each candidate is assigned a risk tier so you know what's safe to act on.

***Bayes excludes tables it can't safely propose dropping — for example, inbound shared datasets (such as Snowflake Secure Data Shares) whose storage is billed to the provider, not your account.
Storage > Risk Tiers
Every candidate in the Storage optimization tab carries a risk tier. The tier tells you how much review a table is likely to need before you act on it — it is a way to order your cleanup queue, not a verdict on whether a table can be deleted.
Newly connected warehouses - A warehouse needs query log activity on at least 7 of the last 30 days before its tables appear as candidates.
Tiers are built from signals Monte Carlo can observe:
- Downstream dependencies — how many assets depend on the table, from lineage
- Read and write activity — whether the table was read or written during the lookback window
- Staleness — how long since the table last changed
- Location — whether the table sits in a production path or elsewhere
| Tier | Typically indicates | What to do |
|---|---|---|
| Low risk* | No downstream dependencies, no observed reads, no recent changes. | Start here. These need the least investigation. |
| Needs review | Mixed signals — for example no reads but downstream dependencies present, or recent write activity. | Check with the table's owner before acting. |
| High risk | Downstream dependencies, recent activity, or a production path. | Not a cleanup target. Investigate only if you have specific reason to. |
Filter the tab by risk level to work through one tier at a time.
*Low risk does not mean safe to delete immediately - Monte Carlo can see reads, writes, lineage and change history. It cannot see retention requirements, regulatory obligations, disaster recovery copies, consumers outside the warehouse, or your team’s plans for a table.
Compute Analysis
Bayes analyzes query groups (recurring query patterns) across your warehouse to surface runtime and performance opportunities. It can:
- Rank query groups by runtime, duration, credit cost, or frequency
- Identify runtime anomalies and outliers (e.g. a job whose max runtime is many times its average)
- Surface anti-patterns — full scans where a partition filter would help, high-frequency micro-batches, redundant duplicate pipelines
- Profile a query group's performance over time
Warehouse Usage Analysis
For Snowflake environments, Bayes adds warehouse-level analysis:
- Ranks warehouses by credit consumption
- Identifies over-provisioned or under-utilized warehouses
- Suggests right-sizing (which can mean up-sizing — a larger warehouse that finishes faster sometimes costs less overall)

Scheduled digests
Rather than re-running an investigation each time, subscribe to a scheduled digest — a recurring summary of your storage and query-group findings, delivered by email on a cadence you choose. Open the Scheduled digests tab and set:
- Warehouse — which connection to report on
- Audience — the recipients (any Monte Carlo audience - email or slack)
- Cadence — how often it sends, and on which day
Digests are a low-effort way to keep the team aware of cost trends and new opportunities without anyone logging in to look.

Accessing Cost & Performance
In the Monte Carlo UI
Open Cost & Performance in the main navigation.
From any finding you can expand the full detail, copy the suggested change, create a ticket to track the work, or Export the full list to act on it outside the product. You can also click "Find out more" on any finding to open it in a conversation with Bayes, with full context preloaded — from there, ask any follow-up question to investigate deeper.
Via the Agent Toolkit (MCP)
The Agent Toolkit exposes Bayes' cost and performance analysis to AI coding tools like Claude Code and Cursor. The storage-cost-analysis skill covers storage waste; the performance-diagnosis skill covers query and runtime analysis. For example:
- "What's costing us the most in storage?"
- "Identify unread tables in my warehouse"
- "Estimate cleanup savings for unused tables"
Configuring Cost & Performance Insights
You control cost and performance analysis from account-wide settings, including the scope Bayes analyzes.
Account-wide settings
Go to Settings > Cost & performance agent
Usage — Automatically analyze warehouse usage to surface cost and performance opportunities. Choose by:
- Unlimited (the default)
- Limited / free, which uses no credits and shows one finding per weekly run: the one with the highest estimated savings
Warehouses to analyze — Choose which warehouses Bayes can analyze:
- All warehouses — every supported warehouse in the account
- Target specific warehouses — select individual warehouses from the dropdown
Findings — Choose which types of saving opportunities Bayes surfaces:
- All (the default)
- Storage only — unused and stale tables, storage optimization candidates
- Compute only — warehouse sizing, idle time, and query optimization
Security & Data Privacy
When Bayes analyzes cost and performance, it uses only metadata, lineage, and query-log data — it does not access or process the actual data inside your tables, and it never runs queries against your warehouse to perform this analysis. For more detail, see AI Features and Technical Information.
FAQ
Does Bayes change anything in my warehouse?
No. It proposes and provides the change to make; you decide what to act on. Monte Carlo is read-only by design.
How does it know a table is safe to drop?
It combines two signals: no read activity in the query logs, and no downstream dependencies in lineage. Because that's based on captured activity, Bayes assigns a risk tier and suggests confirming before dropping.
Why don't I see dollar estimates for my warehouse?
Dollar estimates require a clear cost-to-usage mapping, which is not always available. For a couple of platforms Bayes shows table sizes and runtime.
Can it analyze object storage (S3 / GCS buckets) directly?
Not today — the analysis is table-centric, working from catalogued warehouse tables.
Why are there no storage candidates for a warehouse I just connected?
A warehouse must be connected for at least 7 days before storage candidates appear for it, so there's enough query history to judge. After that, candidates are recalculated several times a day. The agent's findings are produced once a week, so a new finding can take up to a week to appear.
Can I see usage for each user?
No. Usage is measured for the whole account, from daily query-hours and the total storage in the warehouses in scope. See the consumption rates. To use less on the Unlimited tier, limit the agent to specific warehouses, or set Findings to Storage only or Compute only, in Settings > Agentic operations > Cost & Performance agent.
Updated 4 days ago
