Skip to content

Overview

Altimate AI's Auto Tune agents automatically optimize your warehouse costs by dynamically adjusting key configurations based on real-time workload demands. Our system intelligently manages:

  • Auto-suspend – Automatically pauses idle warehouses to eliminate unnecessary costs
  • Max cluster count – Scales compute resources up or down to match query demand

High-Level Approach to Calculating Savings

Core Principles

  • Conservative: Underestimates when uncertain
  • Transparent: Provides min-max ranges reflecting real variability and overall confidence
  • Auditable: Every calculation traces to specific events
  • Performance-Aware: Accounts for potential speed/cost trade-offs

Our savings calculation methodology based on whether Auto Tune is enabled:

1. Estimated Savings (Auto Tune Off)

For warehouses without Auto Tune enabled, we replay historical workloads to project potential savings:

  • Collect warehouse metrics from previous months
  • Replay historical workloads through our optimization model to determine what decisions it would have made
  • Compare our replayed optimization decisions against Snowflake's actual default behavior during that period
  • Calculate the estimated savings range based on the difference between optimized and actual events.

2. Realized Savings (Auto Tune On)

For warehouses with Auto Tune actively managing configurations:

  • Capture real-time metrics reflecting our optimization decisions
  • Estimate Snowflake's default behavior to establish a baseline
  • Compare actual optimized costs against the baseline scenario
  • Calculate the realized savings range achieved through optimization

Warehouse Suspension

Purpose: Reduces idle warehouse costs by intelligently predicting when warehouses can be safely suspended, eliminating wasted runtime between queries without impacting performance.

Estimated Savings (Auto Tune - Off)

  1. Reconstruct warehouse timeline: Map all query executions during active periods for every warehouse
  2. Determine eligible savings periods: Replay the model across the timeline to identify when would Altimate would adjust and try to suspend the warehouse due to it being idle
  3. Calculate credits: Calculate the time difference between where we would suspend to when snowflake suspended and that is the savings we have achieved. That would be converted to credits and cost based on the warehouse size and warehouse type.

Example

  • Warehouse: Medium (4 credits/hr), 60-second auto-suspend
  • Runtime: Warehouse was on for 6:00 AM – 6:10 AM
  • Opportunity: We detected that the warehouse would be idle at 6:09 AM (1 minute early) and would have suspended then.
  • Savings: 1 min ÷ 60 × 4 credits = 0.067 credits

Realized Savings (Auto Tune - On)

  1. Capture Auto Tune suspension event: Record the exact timestamp when Altimate suspends the warehouse based on the model prediction

  2. Establish baseline comparison: Determine when Snowflake's default auto-suspend would have triggered based on the configured threshold (e.g., 10-minute auto-suspend setting)

  3. Measure actual idle duration: Track how long the warehouse remained suspended before the next query arrived

  4. Calculate actual savings window: The savings window is the minimum of -

  5. Time between Auto Tune suspension and default auto-suspend threshold, OR

  6. Time between Auto Tune suspension and when the next query actually arrived This ensures we only count savings for idle time that was actually eliminated, capped by either the default auto_suspend or when the warehouse was restarted.

  7. Convert to credit savings: Multiply the saved idle time by the warehouse's credit consumption rate (based on warehouse size and type)

Example 1: Full savings (no query interruption)

  • Warehouse: Medium (4 credits/hr) with 10-minute default auto-suspend
  • Auto Tune suspended at: 10:00 AM
  • Next query arrived at: 10:15 AM
  • Snowflake would have suspended at: 10:10 AM
  • Savings window: 10 minutes (capped by default threshold)
  • Credits saved: 10 min ÷ 60 × 4 credits/hr = 0.67 credits

Example 2: Partial savings (query arrives early)

  • Warehouse: Medium (4 credits/hr) with 10-minute default auto-suspend
  • Auto Tune suspended at: 10:00 AM
  • Next query arrived at: 10:05 AM (warehouse resumes)
  • Snowflake would have suspended at: 10:10 AM (never reached)
  • Savings window: 5 minutes (capped by next query arrival)
  • Credits saved: 5 min ÷ 60 × 4 credits/hr = 0.33 credits

Warehouse Scaling

Purpose: Optimizes cluster counts in multi-cluster warehouses by scaling down idle clusters more aggressively than Snowflake's native scaling policies, eliminating wasted compute spend during periods of reduced query demand.

Estimated Savings (Auto Tune - Off)

  1. Reconstruct cluster utilization timeline: Map all query executions and warehouse cluster timelines.
  2. Replay Auto Tune scaling logic: Simulate when Altimate would have scaled down each cluster based on our optimization model
  3. Calculate savings window: For each cluster's idle period, determine the actual time saved by comparing:
  4. The savings window is the minimum of these two durations
    • we can only save time up until either work resumes
    • or time till the cluster was suspended by snowflake.
  5. Convert to credit savings range: Multiply the saved idle time by each cluster's credit consumption rate (based on warehouse size and type), then sum across all clusters. Apply a 10% range to account for variability, providing minimum and maximum savings estimates.

Realized Savings (Auto Tune - On)

  1. Capture Auto Tune scale-down event: Record when Altimate scales down a cluster, noting the exact timestamp and which cluster was removed
  2. Measure time until next query: Track the duration from scale-down until the next query arrives on that cluster (indicating capacity was needed again)

  3. Calculate credit savings range: Both are capped by the time to next query because we can only save time until capacity is needed again. We calculate a savings range as we are unsure when snowflake would turn it off. Based on past documentation and observation - A Multi cluster warehouse would do 2-3 checks on 1 minute interval

  4. Minimum savings: min(2 minutes, time to next query)

  5. Maximum savings: min(3 minutes, time to next query)

Example 1: Next query arrives within the maximum threshold

  • Warehouse: Large (8 credits/hr per cluster)
  • Auto Tune scaled down at: 10:00 AM (removed 1 cluster)
  • Next query arrived at: 10:02:30 AM (2.5 minutes later)
  • Calculation:
    • Minimum: MIN(2 min, 2.5 min) = 2 min
    • Maximum: MIN(3 min, 2.5 min) = 2.5 min
  • Credits saved:
    • Minimum: 2 min ÷ 60 × 8 credits/hr = 0.27 credits
    • Maximum: 2.5 min ÷ 60 × 8 credits/hr = 0.33 credits

Example 2: Next query arrives before minimum threshold

  • Warehouse: Large (8 credits/hr per cluster)
  • Auto Tune scaled down at: 10:00 AM (removed 1 cluster)
  • Next query arrived at: 10:01:30 AM (1.5 minutes later)
  • Calculation:
    • Minimum: MIN(2 min, 1.5 min) = 1.5 min (capped by when query arrived)
    • Maximum: MIN(3 min, 1.5 min) = 1.5 min (also capped by when query arrived)
  • Credits saved:
    • Minimum: 1.5 min ÷ 60 × 8 credits/hr = 0.20 credits
    • Maximum: 1.5 min ÷ 60 × 8 credits/hr = 0.20 credits

Example 3: Next query arrives beyond maximum threshold

  • Warehouse: Medium (4 credits/hr per cluster)
  • Auto Tune scaled down at: 2:00 PM (removed 1 cluster)
  • Next query arrived at: 2:05 PM (5 minutes later)
  • Calculation:
    • Minimum: MIN(2 min, 5 min) = 2 min (full minimum threshold)
    • Maximum: MIN(3 min, 5 min) = 3 min (full maximum threshold)
  • Credits saved:
    • Minimum: 2 min ÷ 60 × 4 credits/hr = 0.13 credits
    • Maximum: 3 min ÷ 60 × 4 credits/hr = 0.20 credits

Total Annualized Potential Cost Savings

We calculate your total potential annual autonomous savings by aggregating all optimization types (suspension and scaling) across your warehouses over a rolling 28-day period. For each warehouse on each day, we use estimated savings if Auto Tune was off (projected savings) or realized savings if Auto Tune was on (actual savings). These daily totals are summed to establish your 28-day baseline, which we then annualize by multiplying by 365 / 28.