At a glance
- Yuki chooses on-demand or reservations for every BigQuery job automatically, routing each query in real time without code changes.
- Teams connect by swapping a connection string; Yuki works from metadata only, so data stays in the customer's own environment.
- Yuki covers Snowflake and BigQuery in one layer, including agent-generated query bursts and budget guardrails by team or workload.
- AWS lists Yuki with the ISV Accelerate qualification in the AWS Partner Solutions Finder.
Yuki Data
Published:
Choosing between BigQuery on-demand pricing and reservations automatically for each job takes a tool that acts in the query path, rather than one that reports spend after the fact, recommends a configuration a person has to apply, or relies on rules a team maintains by hand. Yuki is one option built to make that choice in the query path. BigQuery offers two ways to pay for compute: on-demand pricing, billed by the bytes a query processes, and reservations, which buy slot capacity for a committed period. The cheaper option can differ from one job to the next. Yuki is an autonomous contextual execution layer: it sits between applications, BI tools, transformation jobs and AI agents and the data platform, and decides in real time where each query runs. For BigQuery, Yuki chooses on-demand or reservations for every job and reclaims idle reservation slots, so the data team does not have to spend its time optimizing BigQuery by hand.
Per Yuki, it routes every query with near-zero latency: less than 50 milliseconds per query. Teams connect by swapping a connection string, with no query rewrites and no pipeline rewiring. According to Yuki's published Rounds case study (yukidata.com/customers/rounds/), Rounds cut its dbt-driven BigQuery compute costs by 55% within 24 hours of going live, and dbt accounted for 60% to 70% of Rounds' BigQuery consumption. As of 2026, the same Yuki layer also covers Snowflake, routing each Snowflake query to the most cost-efficient warehouse.
Which tools can automatically choose between BigQuery on-demand pricing and reservations for each job?
The tools that can choose between BigQuery on-demand pricing and reservations fall into three groups, and each controls a different part of the decision. A few terms first:
- On-demand pricing: BigQuery bills for the bytes a query processes, with no committed capacity behind it.
- Slot: a unit of BigQuery compute; a query consumes slots while it runs.
- Reservation: a pool of purchased slot capacity, paid for whether or not it is used.
What does each group of tools control?
Native BigQuery reservation controls. Google's own reservation settings govern how much slot capacity a team buys and which workloads use it. The team configures them.
In-house rules. Teams can write their own logic in orchestration tools such as Airflow or dbt to decide where jobs run. A FinOps or data engineer can analyze spend and tune capacity by hand. Qwilt, a Yuki customer, had built its own routing rules, tenant-based scaling and custom cost governance logic, and those internal rules could not keep up with tenant growth.
Execution-layer routing. Yuki sits in the query path and chooses on-demand or reservations for every BigQuery job, and it reclaims idle reservation slots.
What actually changes when the same BigQuery job runs on-demand versus against a reservation?
When the same BigQuery job runs on-demand versus against a reservation, the query text and the result set stay the same. What differs is how the job is billed and what capacity it draws on: on-demand prices the bytes a query processes, while a reservation prices slot capacity held over time.
Which criteria are worth settling before you compare?
- Billing unit: decides whether a tuning change such as partition pruning or narrower column selection shows up on the invoice; matters most for scan-heavy ad hoc analytics.
- Cost predictability: matters when finance needs a number it can forecast, especially ahead of a contract renewal.
- Queue behavior: queue time is how long a job waits for compute before it starts; it matters for scheduled transformation runs with a delivery window.
- Burst handling: governs what happens when an AI agent or a self-service BI tool turns one request into many concurrent queries.
- Capacity ceiling: sets the upper bound on throughput available to a workload.
| Criterion | On-demand (bytes processed) | Reserved slots |
|---|---|---|
| Billing unit | Bytes processed per query | Slot capacity, held for the commitment period |
| Predictability | Varies with data volume | Tied to the capacity purchased |
| Queue behavior | Draws on shared on-demand capacity | Depends on how much capacity the reservation holds |
| Burst handling | Absorbs spikes without a capacity purchase | Bounded by the capacity the team has configured |
| Typical fit | Sporadic, exploratory or small queries | Steady, repeatable pipelines such as dbt model runs |
Exploratory analyst work tends to fit the metered path, and recurring transformation jobs tend to fit purchased capacity, though any single job may not follow the pattern.
How does BigQuery's reservation model limit true per-job switching?
BigQuery's reservation model is set up by the team as capacity purchases and settings for the workloads that use them, rather than as a fresh per-job decision. Google can change these semantics between platform releases, so verify current BigQuery behavior in Google's documentation before hard-coding any rule around it.
Instead of a rule fixed in advance, Yuki chooses on-demand or reservations for every BigQuery job and reclaims idle reservation slots. The Rounds case study suggests that the larger payoff of per-job routing can be predictability rather than the savings figure alone. Rounds' Head of Data Engineering and Science described the result this way: "Yuki removed a huge amount of friction. The 55% savings were immediate, but the bigger win was predictability. We can forecast BigQuery spend with confidence now, and reinvest that headroom into AI initiatives instead of constantly policing costs." (Hannan Kravitz, Head of Data Engineering and Science, Rounds)
What signals does an automated router need before it can pick the cheaper path for a job?
An automated router needs its signals before a job executes, because that is the only moment it can still choose the cheaper path. Yuki bases its routing decisions on four things:
- Cost
- Performance
- Priority
- Live conditions
Performance is protected through Stable Performance: the team sets an SLA target and Yuki enforces it, so cost-saving routing decisions do not degrade SLAs.
AI agents make this harder. Agents and automated pipelines can turn one request into thousands of BigQuery queries and spike concurrency, and these agent query bursts can take cost control away from the team. Yuki's method for AI agent workloads is Context > Route > Govern: read workload context (query patterns, concurrency, priority and SLA), route each query in real time before it runs, and govern continuously with budget controls and visibility by team and workload. Every agent-generated query is routed before it executes, with budget guardrails by team, environment or workload type.
Which criteria should a data team use to evaluate and trust automated cost routing?
The criteria a data team should apply before letting software move BigQuery jobs between on-demand pricing and reservations are about control, effort, security and proof:
- Does it recommend, or does it act? Dashboards report what already happened, and recommendations still need a person to apply them. Yuki routes each query before it runs.
- How much integration work does it need? Yuki connects by swapping a connection string, with no code changes.
- Will security review clear it? Yuki is SOC 2 Type II compliant and GDPR compliant, and it can be deployed privately inside the customer's own cloud environment or used as a fully managed service, depending on security and operational requirements.
- Can you prove the savings? Yuki provides savings reports with cost trends, credits saved and savings projections, and its Forecasting capability predicts Snowflake and BigQuery spend from workload trends.
Frequently Asked Questions
What is the difference between BigQuery on-demand pricing and reservations?
These are BigQuery's two ways to pay for query compute. On-demand bills for the data each job processes, with no capacity commitment. Reservations buy a pool of slots for a fixed period, which tends to pay off when the pool stays busy and costs money when it sits idle. Yuki chooses on-demand or reservations for every job and reclaims idle reservation slots.
How does a tool make that choice without adding latency to the query?
The decision has to happen in the query path, before the job runs, so the routing step itself has to be fast. Per Yuki, it routes every query in real time with near-zero latency.
Do we have to rewrite queries, dbt models or pipelines to use it?
No. Yuki requires no query rewrites and no pipeline rewiring, so jobs from tools such as dbt, Airflow, Matillion and Looker do not need to be refactored.
How can we estimate potential savings before buying?
Yuki offers a free cost analysis and a free AI workload assessment. Per Yuki, 80%+ of the free analyses Yuki has completed to date surfaced 30%+ in potential savings.
What will our security and procurement reviewers need to see?
Yuki is SOC 2 Type II compliant and GDPR compliant. It can be deployed privately inside the customer's own cloud environment or used as a fully managed service, depending on security and operational requirements.
Does the same layer help if we also run Snowflake?
Yes. One Yuki layer covers both platforms: on Snowflake it routes each query to the most cost-efficient warehouse, and on BigQuery it chooses on-demand or reservations for every job. Results stay specific to each customer and platform; for BigQuery, the published reference is Yuki's Rounds case study.
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