Manage warehouse compute

​

Overview

When you connect Hightouch to your warehouse as a source, Hightouch runs directly in that warehouse. When you build audiences, run syncs, calculate traits, or execute journeys in Hightouch, your warehouse runs SQL queries to power those actions. Because of this, Hightouch contributes to your warehouse compute usage. Many customers find that Hightouch accounts for less than 1% of their annual warehouse compute costs. This guide explains how to plan for, monitor, and optimize that usage intentionally. You’ll learn:

• What typically increases warehouse cost
• How your data structure affects performance
• What you can change in your warehouse
• What you can change in Hightouch

​

How warehouse billing works

Most warehouses charge separately for:

• Compute – the processing power used to run queries
• Storage – how much data is stored
• Platform services – management and metadata features

Hightouch primarily affects compute. Each time your warehouse evaluates an audience, calculates a trait, runs a sync, or processes a journey step, compute is consumed.

​

What usually increases warehouse compute

Warehouse cost typically goes up for one of five reasons:

1. Queries run more often
2. Queries do more work
3. More queries run at the same time
4. The warehouse is not the right size for the workload
5. The warehouse stays on longer than necessary

​

1. Queries run more often

More frequent queries increase total compute usage. Common drivers include:

• Syncs scheduled more frequently
• Triggered journeys running continuously
• Audiences recalculating often

If a use case doesn’t require real-time updates, consider lowering the schedule. Many background jobs work just as well daily or weekly.

​

2. Queries do more work

Queries that perform large joins, heavy aggregations, or full-table scans consume more compute. Queries become more expensive when they:

• Join large tables repeatedly
• Scan long time ranges of event data
• Perform heavy calculations each time they run

To reduce compute, move complex joins or calculations earlier in your data pipeline (for example, in dbt) and have Hightouch read from a simpler, materialized table. This avoids repeating the same heavy work every time a sync, audience, or trait runs.

​

3. More queries run at the same time

Concurrency (sometimes called parallelization) refers to how many queries run at the same time. When many syncs, journeys, or recalculations start at once, your warehouse must process them simultaneously. Higher concurrency can:

• Increase short-term cost spikes
• Slow down interactive work

Adjust concurrency based on what your warehouse can handle and how quickly jobs need to finish. If jobs are not time-sensitive, spreading them out can reduce peak usage.

​

4. The warehouse is not the right size for the workload

Warehouse size (sometimes called a compute engine) determines how much compute is available at once.

• Larger warehouses process more data per unit time but cost more while running.
• Smaller warehouses cost less while running but may introduce wait times during spikes.

Start with a modest warehouse size and scale up only if jobs regularly take too long. Many customers begin with a Medium warehouse.

​

5. The warehouse stays on while idle

Warehouses that remain running when no queries are executing still incur compute cost. Idle behavior depends on:

• Auto-suspend (how quickly the warehouse turns off when it’s idle)
• How jobs are scheduled throughout the day
• Whether background jobs run continuously or in batches

Many environments default to ~10 minutes of inactivity. Lowering auto-suspend to around 1 minute—while keeping auto-resume enabled—is often one of the simplest ways to reduce unnecessary compute usage.

​

How Hightouch uses your warehouse

Understanding how Hightouch workloads behave helps explain why certain cost drivers appear in practice. Hightouch workloads fall into a few broad categories. Understanding them helps you decide what needs to be fast and what can run in the background.

​

Interactive workloads

Interactive workloads are user-driven and show up directly in the platform UI:

• Audience previews in Audiences
• Trait previews in Traits
• Insights and dashboards in Intelligence
• Identity Resolution jobs
• Some AI-assisted workflows

These workloads prioritize quick load times because users expect fast responses.

​

Batch workloads

Batch workloads run on a schedule or in the background:

• Sync runs to downstream tools
• Large trait recalculations
• Audience snapshots and backfills

These workloads prioritize processing large volumes of data reliably and can often run on a smaller or lower-priority warehouse.

​

Background jobs

Some jobs run periodically but are not time-sensitive:

• Match Booster enrichment refreshes
• Cached column suggestion refreshes
• Non-urgent audience updates

Because these jobs tolerate delay, they are good candidates for:

• Lower-frequency schedules
• Smaller, dedicated warehouses
• Running during off-peak windows

​

Strategies to reduce warehouse compute

There are three main levers that affect how much compute Hightouch consumes:

1. Warehouse configuration
2. Data modeling and table design
3. Hightouch configuration

Optimizing across all three provides predictable performance and cost control.

​

1. Warehouse configuration

​

Use a dedicated warehouse and user for Hightouch

Create:

1. A dedicated warehouse for Hightouch
   • Allows you to clearly see what Hightouch costs and prevents it from competing with other workloads.
2. A dedicated user (or service account) and role with access only to the data Hightouch needs
   • Makes it easier to control access and audit the queries Hightouch runs.

Starting here provides the clearest cost visibility and strongest governance controls.

​

Choose the right warehouse size

Start with a modest size (often Medium) and scale based on runtime and queueing patterns. Increase size if jobs regularly exceed acceptable runtimes.
Decrease size if the warehouse is mostly idle.

​

Use serverless or multi-cluster options when appropriate

If your warehouse supports dynamic scaling (for example, serverless or multi-cluster configurations), consider enabling it. These options:

• Add capacity automatically during busy periods
• Scale down when demand drops
• Prevent you from permanently running a larger warehouse

This is especially helpful if you have steady daily usage with occasional spikes.

​

Reduce idle time

Lower auto-suspend (around 1 minute is common) and keep auto-resume enabled to prevent idle compute waste.

​

Control how many queries run at once

Adjust how many queries your warehouse can run at the same time.

• Increasing concurrency allows more work to finish sooner but may increase peak usage.
• Decreasing concurrency smooths load but can extend total runtimes.

Adjust this setting based on whether cost spikes or missed deadlines are the bigger concern.

​

Set query time limits

Configure maximum query runtimes to prevent a single query from running indefinitely.

• Set reasonable time limits to guard against runaway compute usage.
• If queries frequently hit the limit, optimize the underlying model before increasing the threshold.

​

Use budgets and resource monitors

Most warehouses provide built-in tools to:

• Track credit usage
• Set spending thresholds
• Alert on anomalies

Using these with a dedicated Hightouch warehouse makes unusual usage easier to detect and investigate.

​

Route different workloads to different warehouses

If your warehouse supports multiple compute engines, route:

• Interactive workloads (e.g., audience previews, dashboards, IDR, AI-powered workflows) to a warehouse tuned for low-latency (fast UI queries)
• Batch workloads (e.g., syncs, heavy traits) to a warehouse tuned for throughput (large scheduled jobs)

This keeps the UI responsive during heavy batch windows and separates cost management by workload type.

---

​

2. Data modeling and table design

These strategies reduce how much work your warehouse performs each time Hightouch runs a query.

​

Turn complex queries into tables

If a model performs heavy joins or aggregations every time it runs, your warehouse must repeat that work for every sync, audience, or trait. This can quietly increase compute usage over time. To reduce repeated work:

• Run those joins and calculations earlier in your data pipeline (for example, in dbt) and have Hightouch read from a simpler, materialized table or view that serves as a presentation layer.
• Simplify the model Hightouch reads from.

If a simple SELECT * FROM your_model LIMIT 1 query runs slowly, the model is likely doing too much work at read time and should be refactored.

​

Structure your tables intentionally

Hightouch frequently runs queries that filter, join, and aggregate data to compute audiences, traits, and journey steps. Efficient table structure reduces repeated work and lowers compute usage. A practical default is:

• Keep core warehouse layers normalized by default (separate entities into separate tables and follow good modeling practices).
• Expose a flattened, presentation-layer table or view for Hightouch when it clearly removes slow, repeated joins for the way you actually query the data (especially for Customer Studio parent models).
• Avoid denormalizing very large fact tables unnecessarily, as this can increase how much data the warehouse must read; instead, denormalize surgically where it meaningfully reduces query cost for common patterns.

For time-series and large tables, consider adding clustering keys (how your warehouse organizes data to make common filters faster) once you understand real query patterns.

​

3. Hightouch configuration

These strategies involve controls inside Hightouch that affect how often and how intensively queries run.

​

Simplify the models that power your Customer Studio schema

Customer Studio uses parent, related, and event models to evaluate audiences and traits. Your schema design directly affects how much work your warehouse must perform each time a query runs. If these models contain complex logic, repeated joins, or large amounts of unnecessary data, every audience preview, sync, and journey step becomes more expensive. As a general rule, use:

• Parent models for users or accounts
• Related models for supporting records (like orders or subscriptions)
• Event models for behavioral activity

Keeping these entities separate by default makes queries more predictable and reduces repeated work.

​

Limit large event tables to what activation requires

Event tables can grow quickly. To keep them efficient:

• Limit time windows to what activation actually requires
• Use pre-aggregated tables for common metrics when helpful
• Organize tables based on the columns most often used in filters (for example, date or tenant)

For large event models, partition event tables by a timestamp column and always include a partition-aligned time filter in audiences and traits to avoid full-table scans. Traits can perform aggregations directly for lighter or audience-specific metrics. For heavy or widely reused metrics, pre-aggregated tables or materialized views are usually more efficient than re-aggregating on every trait or audience run.

​

Schedule background jobs at the lowest effective frequency

Many workloads do not require high-frequency updates.

• Reduce sync frequencies for enrichment updates, cached column suggestions (often daily or weekly is sufficient), and non-time-sensitive audiences
• Choose schedules that match when stakeholders actually use the data

See Schedule syncs in the UI to learn how to set up sync schedules.

​

Use sampling and fast queries for exploration

Audience and trait exploration does not always require exact counts.

• Sampling maintains a smaller, regularly refreshed copy of your parent model.
• Fast queries instruct Hightouch to use that sample for previews instead of scanning the full dataset.

This significantly reduces the amount of data scanned during exploration and provides much faster previews in exchange for approximate (rather than exact) counts. Fast queries and sampling affect previews only. They do not change final audience membership or sync volume.

​

Enable fast queries in audiences

Use fast queries when:

• Brainstorming logic
• Validating filters
• Comparing audience shape

When you need exact numbers (for example, right before go-live, for small or highly targeted audiences, or for compliance and reporting) turn fast queries off to run against the full dataset. To enable fast queries:

1. Go to Customer Studio → Audiences.
2. Open an audience.
3. Enable the Fast queries toggle in the audience header, if available.

​

Enable sampling in models

Sampling is especially helpful when:

• Your parent model contains millions or billions of rows
• Audience previews feel slow during filter changes
• Data teams want to reduce compute usage from iterative exploration

See Sampling for detailed setup instructions and configuration guidance.

​

Control how many jobs Hightouch runs at once

Hightouch can run many jobs at the same time. This setting controls how many Hightouch syncs and background jobs are initiated in parallel. It is separate from warehouse-level concurrency.

• Higher concurrency reduces how long jobs take overall but increases short-term compute usage.
• Lower concurrency spreads work out over time and reduces peak load.

In general, keep Hightouch workspace concurrency at or below your warehouse’s effective concurrency limits (and any queue timeouts) so sync queries aren’t killed while waiting for a warehouse slot. Match this setting to your warehouse limits and job timing requirements.

---

​

If warehouse costs increase unexpectedly

Start with these checks:

1. Did sync or journey frequency change?
2. Did event volume grow?
3. Did new audiences or traits launch?
4. Did concurrency settings change?
5. Did auto-suspend settings change?
6. Did a data model change introduce heavier joins or scans?

Most cost increases can be traced to one of these changes.