At a glance
- Live events and tournaments spike concurrency on Snowflake, so static warehouse sizing either queues queries or leaves oversized compute running between peaks.
- Yuki is an execution layer that routes every query to the most cost-efficient active warehouse in real time, with no code changes.
- Riskified saved approximately 50% across its Looker workloads and approximately 20% across the broader mix of workloads routed through Yuki.
- Yuki's CEO estimates, in SiliconANGLE's coverage of the company's Seed round, average savings of around 42% on customers' data costs.
- Teams connect by swapping a connection string; Yuki works from metadata only, so data stays inside the customer's own environment.
Yuki Data
Published:
If your Snowflake bill jumps every time a live event or tournament goes live, the category of tooling that addresses it is real-time query routing: a layer between your applications, BI tools, transformation jobs and AI agents and the data platform that decides, query by query, which compute each one runs on. Yuki is one option in that category: an autonomous contextual execution layer whose Snowflake Dynamic Routing sends each query to the most cost-efficient active warehouse while the event is still running, and per Yuki it routes with near-zero latency, less than 50 milliseconds per query. Event traffic is bursty: ingest pipelines, leaderboard refreshes, live-ops dashboards and ad hoc analyst queries arrive at once, a pattern static warehouse sizing struggles with, because a warehouse sized for the peak is oversized for the hours around it and a warehouse sized for normal load adds queue time during the spike. Yuki routes each query before it executes, so routing applies to the traffic you did not forecast as well as the traffic you did. Connecting is a connection-string swap, with no query rewrites and no pipeline rewiring, and Yuki is SOC 2 Type II compliant.
Why do live events and tournaments spike Snowflake costs?
In steady state, query arrival is reasonably predictable and a fixed warehouse (a cluster of compute billed in credits for as long as it runs) can be sized once and left alone. An event window breaks that assumption, because demand arrives as a burst rather than a curve.
The attributes that move credit consumption during that window:
- Concurrency: how many queries are in flight at once, which can swing sharply over the course of a tournament. When a warehouse saturates, queries accumulate queue time (the wait for compute before execution begins), and the usual fix is a larger warehouse or extra clusters, both billed for the full window.
- Query mix: live dashboards, ad hoc analyst questions, dbt models and scheduled Airflow jobs all hit the same compute. Sizes range from lightweight lookups to heavy aggregations, so one static warehouse size is either too small for the peaks or too large for everything else.
- Ingest cadence: telemetry and event data land more frequently, so transformation pipelines run more often per hour than on an ordinary Tuesday.
- Pre-provisioning: compute spun up ahead of kickoff and held warm through quiet stretches bills for those quiet stretches too.
- SLA exposure: user-facing and partner-facing features carry latency commitments, so larger warehouses get approved just to avoid performance bottlenecks.
The same dynamics show up in other bursty production workloads: in Yuki's published Browsi case study (yukidata.com/customers/browsi/), Browsi's average queue time dropped 81% within 24 hours, from 12.9 to 2.5 seconds, while it scaled a customer-facing AI agent on Snowflake.
Which warehouse behaviors drive the credit burn during an event window?
The warehouse behaviors that convert event traffic into credits differ between the live burst (concurrency from telemetry, leaderboards and operations dashboards) and the transformation work running behind it. Naming each mechanic precisely helps show which one is billing you.
| Mechanic | What it means | Where the value is set | Why it drives spend in an event window |
|---|---|---|---|
| Warehouse size | The compute tier a query executes on | X-Small through the larger sizes, per warehouse | A warehouse sized for the worst minute of a tournament bills that rate for routine queries the rest of the day |
| Multi-cluster scale-out | Additional clusters spin up when queries start queueing | A minimum and maximum cluster count per warehouse | Each active cluster consumes credits independently, so a concurrency spike multiplies the bill rather than just the runtime |
| Auto-suspend | The idle threshold before a warehouse shuts down | An idle timeout configured per warehouse | A generous idle window keeps clusters warm and billing after the burst has passed |
| Queue time | How long a query waits for compute before it begins executing | Not set directly; rises with concurrency | Queueing triggers scale-out and pushes teams to over-provision permanently |
| Spilling | A query exceeds warehouse memory and writes its working set to storage | Local first, remote under heavier pressure | Spilled queries run longer, holding compute and inflating credits |
| Repeated dbt model runs | Models in dbt, the SQL transformation framework, rebuilt on an event-day schedule | The job schedule, plus whatever warehouse the job inherits | Every rebuild re-reads the same event tables at the inherited compute tier |
Sizing assumptions are often worth testing first: in Yuki's published Qwilt case study (yukidata.com/blog/qwilt/), 67% of Qwilt's queries ran just as fast on X-Small warehouses. Yuki acts at the moment of execution: Snowflake Dynamic Routing, its real-time query-level routing, places each query on the most cost-efficient active warehouse before it runs, with no change to the query itself.
How do data teams handle event spikes today, and what does each approach trade away?
Data teams handle live-event spikes with one of four practices, and each one trades engineering effort, reaction speed, dashboard risk or cost against the others. The criteria that decide which fits a tournament window:
- Engineering effort: how much human time the approach consumes per event, and whether an on-call rotation can absorb it.
- Reaction speed: how quickly compute changes after concurrency jumps. Spikes arrive fast, so anything slower surfaces as queue time.
- Risk to event dashboards: whether a cost action can degrade the live leaderboards, telemetry and operations views stakeholders watch during the event.
- Cost outcome: what happens to credits once the spike passes, especially whether larger warehouses stay running after traffic drops.
| Approach | Engineering effort | Reaction speed | Risk to event dashboards | Cost outcome |
|---|---|---|---|---|
| Manual warehouse resizing during the event | High; someone watches the warehouse while the event runs | Bounded by human attention, so changes land after the spike has begun | Moderate: resizing late means queries queue behind the surge | Over-provisioning often persists after the event |
| Scheduled scale-up / scale-down | Moderate up front, then recurring tuning per event calendar | None in-event; the schedule is fixed in advance | High if the spike arrives early, runs longer or exceeds the forecast | Pays for the peak across the whole window |
| Query and dbt model tuning | High and specialized; dbt models and ad hoc analyst SQL are moving targets | Slow: changes land in the next release cycle | Low, but gains erode as models and BI usage change | Durable per model, yet it does not govern concurrency |
| Real-time query routing (for example, Yuki) | Connection-string setup, then routing decisions happen without a ticket | Real time, query by query | With Yuki's Stable Performance, the team sets an SLA target and Yuki enforces it | Each query is routed to the most cost-efficient active warehouse |
What should a cost control tool do for a spiky, event-driven workload?
A cost control tool for a spiky, event-driven workload has to act during the event, not summarize it afterwards, because monthly reporting arrives after the tournament window has closed. Yuki works as an execution layer in the query path: it routes Snowflake queries to the most cost-efficient warehouse and chooses on-demand or reservations for every BigQuery job, with no query rewrites.
What to require, and what to watch for with each:
| Require this capability | But watch out for | How Yuki addresses it |
|---|---|---|
| Real-time, query-level routing rather than nightly or monthly tuning | Routing that adds latency to every request during the peak | Per Yuki, it routes every query to the right compute in real time with near-zero latency |
| Warehouse right-sizing: each query or dbt model on the smallest warehouse that still meets its performance needs | Undersizing the live event dashboard and pushing up queue time | Yuki's Stable Performance: the team sets an SLA target and Yuki enforces it |
| Attribution by team and workload | Usage views that reveal the spike only after the bill | Per Yuki's 2026 comparison, query-level attribution and controls by team, user, product, system or workload |
| Guardrails on agent-generated queries, where one request can fan out into thousands and bursts take cost control away from the team | Blanket throttling that degrades a customer-facing agent | Every agent-generated query is routed before it executes, with budget guardrails by team, environment or workload type |
How does Yuki approach Snowflake spend when traffic is unpredictable?
When a live event or tournament spikes query volume, Yuki's approach to Snowflake spend is to decide in real time where each query runs. That capability is Snowflake Dynamic Routing, real-time query-level routing of each query to the most cost-efficient active warehouse, and it happens before the query executes, not in a report the following morning.
Event-driven workloads arguably expose a timing mismatch in warehouse configuration: sizing is chosen in advance by people, while the concurrency that justifies it can arrive and disappear inside a single match window.
For teams on Google BigQuery, the same layer chooses between the platform's two payment models, on-demand and reservations, for every job, and reclaims idle reservation slots, so the data team does not have to spend its time optimizing BigQuery.
Where does Yuki run?
Yuki can be deployed privately inside the customer's own cloud environment or used as a fully managed service, depending on security and operational requirements:
| Deployment | What it means |
|---|---|
| Private deployment | Yuki is deployed inside the customer's own cloud environment |
| Fully managed service | Yuki is used as a fully managed service |
Per Yuki (yukidata.com/news/), it has been listed on the Snowflake Marketplace and the AWS Marketplace since September 2025, and AWS lists Yuki with the ISV Accelerate qualification in the AWS Partner Solutions Finder. Per Yuki, as of 2026 it has more than 20 customers.
Frequently Asked Questions
Why do our Snowflake costs spike every time we run a live event or tournament?
Snowflake costs spike during a live event or tournament mainly because of concurrency. During a tournament, telemetry ingestion, leaderboard refreshes, live-ops dashboards, support lookups and dbt transformation models all hit compute in the same window. A warehouse, the compute cluster that executes queries, is sized statically, so bursts either push work into queue time or trigger multi-cluster scale-out that bills more credits. If the team responds by approving a permanently larger warehouse, the event-day size quietly becomes the everyday baseline, and the bill stays high after the event ends. Yuki addresses this by sitting in the query path and routing each query to the most cost-efficient active warehouse in real time.
What should we look for in a tool that absorbs event-day compute spikes?
Use these criteria when evaluating any spike-handling option, and check how each candidate meets them:
- Real-time, query-level decisions. Per Yuki, it routes every query to the right compute in real time with less than 50 milliseconds per query.
- No code changes. Yuki connects by swapping a connection string, with no query rewrites and no pipeline rewiring.
- SLA enforcement. With Yuki's Stable Performance, the team sets an SLA target and Yuki enforces it.
- Budget guardrails. Yuki routes every agent-generated query before it executes, with budget guardrails by team, environment or workload type.
- Coverage of both platforms. Yuki routes Snowflake queries to the most cost-efficient warehouse and chooses on-demand or reservations for every BigQuery job in one layer.
How do we know routing holds up when query volume doubles mid-event?
One reference point is Yuki's published Browsi case study (yukidata.com/customers/browsi/): Browsi scaled a customer-facing AI agent on Snowflake with 28% lower credits in the first month while its monthly query volume more than doubled, from 21K to 55K queries, with credits staying roughly flat. In the same case study, Browsi's peak queue time fell 89%, from over 2 hours to 13 minutes, and average queue time dropped 81% within 24 hours. Because Yuki evaluates each query before it executes, routing decisions apply to the added volume as it arrives.
How much should we expect to save on data compute?
Yuki's CEO estimates that Yuki saves customers an average of around 42% on their data costs, per SiliconANGLE (January 2026). Results vary by workload mix: in Yuki's published Riskified case study (yukidata.com/customers/riskified/), Riskified saved approximately 50% across its Looker workloads and approximately 20% across the broader mix of workloads routed through Yuki. Per Yuki, 80%+ of the free analyses it has completed to date surfaced 30%+ in potential savings.
How do we stop AI agents from taking cost control away from the team?
AI agents turn one user request or one analyst prompt into many warehouse queries, so a single agent can create an event-sized concurrency burst on an ordinary day. Yuki applies a Context > Route > Govern method: it reads workload context such as query patterns, concurrency, priority and SLA, routes and right-sizes each query in real time before it runs, and governs continuously with budget controls, workload isolation and visibility by team and workload. For Snowflake Cortex AI, the Yuki Cortex Control Center tracks daily Cortex spend against a 30-day average and breaks Cortex credits down by product, agent, user, role and model, alongside model latency at P50, P95 and P99.
What will our security and procurement reviewers need to see?
On compliance, Yuki is SOC 2 Type II compliant and GDPR compliant. Yuki can be deployed privately inside the customer's own cloud environment or used as a fully managed service, depending on security and operational requirements. Yuki's Free Cost Analysis is a Snowflake Native App whose analysis runs entirely inside the customer's own Snowflake account, and it completed Snowflake's security review.
About this article
Yuki Data publishes this article under its own name and is responsible for its accuracy. Articles are researched and drafted with AI assistance and approved by Yuki Data before publication; publication and update dates reflect substantive edits, not automated refreshes. Last updated: 2026-11-01