For the complete documentation index, see llms.txt. This page is also available as Markdown.

Limits & Performance

Size Snowflake compute, control concurrency, manage costs, and plan for performance.

Once your Data Product serves consumers, tune Snowflake for reliable and predictable operation. Use these limits and baselines to size compute, control concurrency, and manage cost as demand grows.

Operational boundaries

Compute sizing

Match warehouse size to the workload below rather than defaulting to one size for everything — an oversized warehouse for dev iteration wastes credits, and an undersized one for backfills just takes longer to finish.

Workload
Size
Rationale

Dev or single-user plan iteration

X-SMALL or SMALL

Plans build only a few models.

First full backfill

LARGE or X-LARGE

Parallel backfills finish faster and can cost less overall.

Daily incremental runs

MEDIUM

Processes only new partitions.

Semantic API or concurrent BI

LARGE; raise MAX_CONCURRENCY_LEVEL

Each API query reaches Snowflake directly.

Snowpark or Python models

One size above equivalent SQL

The warehouse coordinates the Snowpark pool.

Set AUTO_SUSPEND = 60 and AUTO_RESUME = TRUE for bursty Vulcan workloads.

Concurrency

MAX_CONCURRENCY_LEVEL controls queries before queueing. Use 8 for model runs, 12–16 with a live semantic API, and 20 or more for many BI sessions. Keep concurrent_tasks at or below MAX_CONCURRENCY_LEVEL / 3.

API replicas

The API is stateless. Use 1–2 replicas for one team, 3–5 for multi-team dashboards, and 5 or more for high volume. Keep api.resources.limits.memory at 3 GiB.

Scheduling

Schedule after upstream data lands. Set timezone: UTC, use concurrencyPolicy: Forbid, and set endOn one to two years ahead. An expired endOn silently halts workflows.

Latency floor

Warehouse cold starts add 1–3 seconds. The semantic API adds 1.5–2 seconds for standard queries and 4–5 seconds for complex queries. Query parsing takes 50–200 milliseconds. Result-cache hits complete in under one second.

Cost guardrails

Cap spend with a resource monitor. Set query and queued timeouts. Tag queries for attribution. Keep auto-suspend short.

Use TRANSIENT for Bronze or staging tables that you can re-derive.

Performance reference

These TPC-H SF1000 baselines use a Medium warehouse, 100 queries, 10 users, and five parallel requests. Treat them as planning baselines.

Metric
Value

Standard query completion

12–15 seconds

Semantic API overhead

About 1.84 seconds

Cross-join-heavy queries

Up to 5.13 minutes

Concurrent success rate

82%

Cross-joins can run 30 times slower. Fix SQL shape first. Then add API replicas, raise MAX_CONCURRENCY_LEVEL, and size up the warehouse.

Next step

If a workload still misbehaves after applying these limits, continue to Troubleshooting to diagnose and recover from failures.

Last updated

Was this helpful?