Understand slots

A BigQuery slot is a virtual compute unit used by BigQuery to execute SQL queries, Python code, or other job types. During the execution of a query, BigQuery automatically determines how many slots are used by the query. The number of slots used depends on the amount of data being processed, the complexity of the query, and the number of slots available. In general, access to more slots lets you run more concurrent queries, and your complex queries can run faster.

While all queries use slots, you have two options for how you are charged for usage, the on-demand pricing model or the capacity-based pricing model.

By default, you are charged using the on-demand model. With this model, you are charged for the amount of data processed (measured in TiB) by each query. Projects using the on-demand model are subject to per-project and per-organization slot limits with transient burst capability. Most users on the on-demand model find the slot capacity limits more than sufficient. However, depending on your workload, access to more slots may improve query performance. To check how many slots your account uses, see Monitor health, resource utilization, and jobs.

With the capacity-based model, you pay for the slot capacity allocated for your queries over time. This model gives you explicit control over total slot capacity. You explicitly choose the amount of slots to use through a reservation. You can specify the number of slots in a reservation as a baseline amount which is always allocated, or as an autoscaled amount, which is allocated when needed. Reservations with autoscaling slots scale their capacity to accommodate your workload demands. BigQuery allocates slots as workloads change. Reservations with autoscaling slots are only available with BigQuery editions.

Query execution using slots

When BigQuery executes a query job, it converts the SQL statement into an execution plan, comprised of a series of query stages. Stages are in turn comprised of sets of execution steps. BigQuery uses a distributed parallel architecture to run queries. Stages model the units of work that can be executed in parallel. Data is passed between stages by using a distributed shuffle architecture, which is discussed in more detail in this Google Cloud blog post.

BigQuery query execution is dynamic. A query plan can be modified while the query is being processed. Work distribution can be optimized for data distribution as stages are added. In addition, capacity for a query's execution can change as other queries start or finish, or as the autoscaler adds slots to a reservation.

BigQuery can run multiple stages concurrently, can use speculative execution to accelerate a query, and can dynamically repartition a stage to achieve optimal parallelization.

Slot resource economy

If a query requests more slots than are available, BigQuery queues up individual units of work and waits for slots to become available. As progress on query execution is made, and as slots free up, these queued up units of work get dynamically picked up for execution.

BigQuery can request any number of slots for a particular stage of a query. The number of slots requested is not related to the amount of capacity you purchase, but rather an indication of the most optimal parallelization factor chosen by BigQuery for that stage. Units of work queue up and get executed as slots become available.

When query demands exceed slots you committed to, you are not charged for additional slots, and you are not charged for additional on-demand rates. Your individual units of work queue up.

For example,

1. A query stage requests 2,000 slots, but only 1,000 are available.
2. BigQuery consumes all 1,000 slots and queues up the other 1,000 slots.
3. Thereafter, if 100 slots finish their work, they dynamically pick up 100 units of work from the 1,000 queued up units of work. 900 units of queued up work remain.
4. Thereafter, if 500 slots finish their work, they dynamically pick up 500 units of work from the 900 queued up units of work. 400 units of queued up work remain.

BigQuery slots queued up if demand exceeds availability

Fair scheduling in BigQuery

BigQuery allocates slot capacity within a single reservation using an algorithm called fair scheduling.

The BigQuery scheduler enforces the equal sharing of slots among projects with running queries within a reservation, and then within jobs of a given project. The scheduler provides eventual fairness. During short periods, some jobs might get a disproportionate share of slots, but the scheduler eventually corrects this. The goal of the scheduler is to find a balance between aggressively evicting running tasks (which results in wasting slot time) and being too lenient (which results in jobs with long running tasks getting a disproportionate share of the slot time).

Fair scheduling ensures that every query has access to all available slots at any time, and capacity is dynamically and automatically re-allocated among active queries as each query's capacity demands change. Queries complete and new queries get submitted for execution under the following conditions:

• Whenever a new query is submitted, capacity is automatically re-allocated across executing queries. Individual units of work can be gracefully paused, resumed, and queued up as more capacity becomes available to each query.
• Whenever a query completes, capacity consumed by that query automatically becomes immediately available for all other queries to use.
• Whenever a query's capacity demands change due to changes in query's dynamic DAG, BigQuery automatically re-evaluates capacity availability for this and all other queries, re-allocating and pausing slots as necessary.

Fair scheduling in BigQuery

Depending on complexity and size, a query might not require all the slots it has the right to, or it may require more. BigQuery dynamically ensures that, given fair scheduling, all slots can be fully used at any point in time.

If an important job consistently needs more slots than it receives from the scheduler, consider creating an additional reservation with the required number slots and assigning the job to that reservation.

As an example of fair scheduling, suppose you have the following reservation configuration:

• Reservation A, which has 1,000 baseline slots with no autoscaling
• Project A and project B, which are assigned to your reservation

Scenario 1: In project A, you run query A (one concurrent query) that requires high slot usage, and in project B you run 20 concurrent queries. Even though there are a total of 21 queries that are using reservation A, the slot distribution is the following:

• Project A receives 500 slots, and query A runs with 500 slots.
• Project B receives 500 slots that are shared among its 20 queries.

Scenario 2: In project A, you run query A (one concurrent query) that requires 100 slots to run, and in project B you run 20 concurrent queries. Since query A doesn't require 50% of the reservation, then the slot distribution is the following:

• Project A receives 100 slots, and query A runs with 100 slots.
• Project B receives 900 slots that are shared among its 20 queries.

Inversely, consider the following reservation configuration:

• Reservation B, which has 1,000 baseline slots with no autoscaling.
• 10 projects, which are all assigned to reservation B.

Assume the 10 projects are running queries that have sufficient slot demand, then each project receives 1/10 of the total reservation slots (or 100 slots), regardless of how many queries are running on each project.

Slot quotas and limits

Slot quotas and limits provide a safeguard for BigQuery. Different pricing models use different slot quota types, as follows:

• On-demand pricing model: You are subject to a per-project and organization slot limit with transient burst capability. Depending on your workloads, access to more slots can improve query performance.

• Capacity-based pricing model: Reservations quotas and limits define the maximum number of slots you can allocate across all reservations in a location. If you use autoscaling, the sum of your maximum reservation sizes cannot exceed this limit. You are only billed for your reservations and commitments, not for the quotas. For information about increasing your slot quota, see Requesting a quota increase.

To check how many slots you are using, see BigQuery monitoring.

Idle slots

At any given time, some slots might be idle. This can include:

• Slot commitments that are not allocated to any reservation baseline.
• Slots that are allocated to a reservation baseline but aren't in use.

Idle slots are not applicable when using the on-demand pricing model.

By default, queries running in a reservation automatically use idle slots from other reservations within the same administration project. BigQuery immediately allocates slots to an assigned reservation when they are needed. Idle slots that were in use by another reservation are quickly preempted. There might be a short time when you see total slot consumption exceed the maximum you specified across all reservations, but you aren't charged for this additional slot usage.

For example, suppose you have the following reservation setup:

• project_a is assigned to reservation_a, which has 500 baseline slots with no autoscaling.
• project_b is assigned to reservation_b, which has 100 baseline slots with no autoscaling.
• Both reservations are in the same administrative project and there are no other projects assigned to these reservations.

You run query_b in project_b. If no query is running in project_a, then query_b has access to the 500 idle slots from reservation_a. While query_b is still running, it may use up to 600 slots: 100 baseline slots plus 500 idle slots.

While query_b is running, suppose you run query_a in project_a that can use 500 slots.

• Since you have 500 baseline slots reserved for project_a, query_a immediately starts and is allocated 500 slots.
• The number of slots allocated to query_b quickly decreases to 100 baseline slots.
• Additional queries run in project_b share those 100 slots. If subsequent queries don't have enough slots to start, then they queue up until running queries complete and slots become available.

In this example, if project_b was assigned to a reservation with no baseline slots or autoscaling, then query_b would have no slots after query_a starts running. BigQuery would pause query_b until idle slots are available or the query times out. Additional queries in project_b would queue up until idle slots are available.

To ensure a reservation only uses its provisioned slots, set ignore_idle_slots to true. Reservations with ignore_idle_slots set to true can, however, share their idle slots with other reservations.

You cannot share idle slots between reservations of different editions. You can share only the baseline slots or committed slots. Autoscaled slots might be temporarily available but are not shareable as idle slots for other reservations because they might scale down.

As long as ignore_idle_slots is false, a reservation can have a slot count of 0 and still have access to unused slots. If you use only the default reservation, toggle off ignore_idle_slots as a best practice. You can then assign a project or folder to that reservation and it will only use idle slots.

Assignments of type ML_EXTERNAL are an exception in that slots used by BigQuery ML external model creation jobs are not preemptible. The slots in a reservation with both ML_EXTERNAL and QUERY assignment types are only available for other query jobs when the slots are not occupied by the ML_EXTERNAL jobs. Moreover, these jobs cannot use idle slots from other reservations.

Reservation-based fairness

Note: You must enable reservation-based fairness before you can create a predictable reservation.

With reservation-based fairness, BigQuery prioritizes and allocates idle slots equally across all reservations within the same admin project, regardless of the number of projects running jobs in each reservation. Each reservation receives a similar share of available capacity in the idle slot pool, and then its slots are distributed fairly within its projects. This feature is only supported with the Enterprise or Enterprise Plus editions.

The following chart shows how idle slots are distributed without reservation-based fairness enabled:

In this chart, idle slots are shared equally across projects.

Without reservation-based fairness enabled, the available idle slots are distributed evenly across the projects within the reservations.

The following chart shows how idle slots are distributed with reservation-based fairness enabled:

In this chart, idle slots are shared equally across reservations, not projects.

With reservation-based fairness enabled, the available idle slots are equally distributed across the reservations.

When you enable reservation-based fairness, review your resource consumption to manage your slot availability and query performance.

Avoid relying solely on idle slots for production workloads with strict time requirements - these jobs should use baseline or autoscaled slots. We recommend using idle slots for lower priority jobs because as the slots can be preempted at any time.

Slot autoscaling

The following section discusses autoscaling slots and how they work with reservations.

Use autoscaling reservations

You don't need to purchase slot commitments before creating autoscaling reservations. Slot commitments provide a discounted rate for consistently used slots but are optional with autoscaling reservations. To create an autoscaling reservation, you assign a reservation a maximum number of slots (the max reservation size). You can identify the maximum number of autoscaling slots by subtracting the max reservation size by any optional baseline slots assigned to the reservation.

When you create autoscaling reservations, consider the following:

• BigQuery scales reservations almost instantly until it has reached the number of slots needed to execute the jobs, or it reaches the maximum number of slots available to the reservation. Slots always autoscale to a multiple of 50.
• Scaling up is based on actual usage, and is rounded up to the nearest 50 slot increment.
• Your autoscaled slots are charged at capacity compute pricing for your associated edition while scaling up. You are charged for the number of scaled slots, not the number of slots used. This charge applies even if the job that causes BigQuery to scale up fails. For this reason, don't use the jobs information schema to match the billing. Instead, see Monitor autoscaling with information schema.
• While the number of slots always scales by multiples of 50, it may scale more than 50 slots within one step. For example, if your workload requires an additional 450 slots, BigQuery can attempt to scale by 450 slots at once to meet the capacity requirement.
• BigQuery scales down when the jobs associated with the reservation no longer need the capacity (subject to a 1 minute minimum).

Any autoscaled capacity is retained for at least 60 seconds. This 60-second period is called the scale-down window. Any new peak in capacity resets the scale-down window, treating the entire capacity level as a new grant. However, if 60 seconds or more have passed since the last capacity increase and there is less demand, the system reduces the capacity without resetting the scale-down window, enabling consecutive decreases without an imposed delay.

For example, if your initial workload capacity scales to 100 slots, the peak is retained for at least 60 seconds. If, during that scale-down window, your workload scales to a new peak of 200 slots, a new scale-down window begins for 60 seconds. If there is no new peak during this scale-down window, your workload begins to scale down at the end of the 60 seconds.

Consider the following detailed example: At 12:00:00, your initial capacity scales to 100 slots and the usage lasts for one second. That peak is retained for at least 60 seconds, beginning at 12:00:00. After the 60 seconds have elapsed (at 12:01:01), if the new usage is 50 slots, BigQuery scales down to 50 slots. If, at 12:01:02, the new usage is 0 slots, BigQuery again scales down immediately to 0 slots. After the scale-down window has ended, BigQuery can scale down multiple times consecutively without requiring a new scale-down window.

To learn how to work with autoscaling, see Work with autoscaling slots.

Using reservations with baseline and autoscaling slots

In addition to specifying the maximum reservation size, you can optionally specify a baseline number of slots per reservation. The baseline is the minimum number of slots that will always be allocated to the reservation, and you will always be charged for them. Autoscaling slots are only added after all of the baseline slots (and idle slots if applicable) are consumed. You can share idle baseline slots in one reservation with other reservations that need capacity.

You can increase the number of baseline slots in a reservation every few minutes. If you want to decrease your baseline slots, you are limited to once an hour if you have recently changed your baseline slot capacity and your baseline slots exceed your committed slots. Otherwise, you can decrease your baseline slots every few minutes.

Baseline and autoscaling slots are intended to provide capacity based on your recent workload. If you anticipate a large workload that is very different from your workloads in the recent past, we recommend increasing your baseline capacity ahead of the event rather than rely on autoscaling slots to cover the workload capacity. If you encounter an issue with increasing your baseline capacity, retry the request after waiting 15 minutes.

If the reservation doesn't have baseline slots or is not configured to borrow idle slots from other reservations, then BigQuery attempts to scale. Otherwise, baseline slots must be fully utilized before scaling.

Reservations use and add slots in the following priority:

1. Baseline slots.
2. Idle slot sharing (if enabled). Reservations can only share idle baseline or committed slots from other reservations that were created with the same edition and the same region.
3. Autoscale slots.

In the following example, slots scale from a specified baseline amount. The etl and dashboard reservations have a baseline size of 700 and 300 slots respectively.

In this example, the etl reservation can scale to 1300 slots (700 baseline slots plus 600 autoscale slots). If the dashboard reservation is not in use, the etl reservation can use the 300 slots from the dashboard reservation if no job is running there, leading to a maximum of 1600 possible slots.

The dashboard reservation can scale to 1100 slots (300 baseline slots plus 800 autoscale slots). If the etl reservation is totally idle, the dashboard reservation can scale to a maximum of 1800 slots (300 baseline slots plus 800 autoscale slots plus 700 idle slots in the etl reservation).

If the etl reservation requires more than 700 baseline slots, which are always available, it attempts to add slots by using the following methods in order:

1. 700 baseline slots.
2. Idle slot sharing with the 300 baseline slots in the dashboard reservation. Your reservation only shares idle baseline slots with other reservations that are created with the same edition.
3. Scaling up 600 additional slots to the maximum reservation size.

Using slot commitments

The following example shows autoscaling slots using capacity commitments.

Like reservation baselines, slot commitments allow you to allocate a fixed number of slots that are available to all reservations. Unlike baseline slots, a commitment cannot be reduced during the term. Slot commitments are optional but can save costs if baseline slots are required for long periods of time. Slot commitments are used to cover baseline slots for your reservations. Any unused slot capacity is then shared as idle slots across other reservations. Slot commitments don't apply to autoscaling slots. To ensure that you receive the discounted rate for your committed slots, make sure that your slot commitments are sufficient to cover your baseline slots.

In this example, you are charged a predefined rate for the capacity commitment slots. You are charged at the autoscaling rate for the number of autoscaling slots after autoscaling activates and reservations are in an upscaled state. For autoscaling rate, you are charged for the number of scaled slots, not the number of slots used.

The following example shows reservations when the number of baseline slots exceeds the number of committed slots.

In this example, there is a total of 1000 baseline slots between the two reservations, 500 from the etl reservation and 500 from the dashboard reservation. However, the commitment only covers 800 slots. In this scenario, the excess slots are charged at the pay as you go (PAYG) rate.

Maximum available slots

You can calculate the maximum number of slots a reservation can use by adding the baseline slots, the maximum number of autoscale slots, and any slots in commitments that were created with the same edition and are not covered by the baseline slots. The example in the previous image is set up as follows:

• A capacity commitment of 1000 annual slots. Those slots are assigned as baseline slots in the etl reservation and the dashboard reservation.
• 700 baseline slots assigned to the etl reservation.
• 300 baseline slots assigned to the dashboard reservation.
• Autoscale slots of 600 for the etl reservation.
• Autoscale slots of 800 for the dashboard reservation.

For the etl reservation, the maximum number of slots possible is equal to the etl baseline slots (700) plus the dashboard baseline slots (300, if all slots are idle) plus the maximum number of autoscale slots (600). So the maximum number of slots the etl reservation could use in this example is 1600. This number exceeds the number in the capacity commitment.

In the following example, the annual commitment exceeds the assigned baseline slots.

In this example, we have:

• A capacity commitment of 1600 annual slots.
• A maximum reservation size of 1500 (including 500 autoscaling slots).
• 1000 baseline slots assigned to the etl reservation.

The maximum number of slots available to the reservation is equal to the baseline slots (1000) plus any committed idle slots not dedicated to the baseline slots (1600 annual slots - 1000 baseline slots = 600) plus the number of autoscaling slots (500). So the maximum potential slots in this reservation is 2100. The autoscaled slots are additional slots above the capacity commitment.

Autoscaling best practices

1. When first using autoscaler, set the number of autoscaling slots to a meaningful number based on past and expected performance. Once the reservation is created, actively monitor the failure rate, performance, and bill and adjust the number of autoscaling slots as needed.

2. Autoscaler has a 1 minute minimum before scaling down so it is important to set the maximum number of autoscaled slots to balance between performance and cost. If the maximum number of autoscale slots is too large and your job can use all the slots to complete a job in seconds, you still incur costs for the maximum slots for the full minute. If you lower your max slots to half the current amount, your reservation is scaled to a lower number and the job can use more slot_seconds during that minute, reducing waste. For help determining your slot requirements, see Monitor job performance. As an alternative approach to determine your slot requirements, see View edition slot recommendations.

3. Slot usage can occasionally exceed the sum of your baseline plus scaled slots. You are not billed for slot usage that is greater than your baseline plus scaled slots.

4. Autoscaler is most efficient for heavy, long-running workloads, such as workloads with multiple concurrent queries. Avoid sending queries one at a time, since each query scales the reservation where it will remain scaled for a 1 minute minimum. If you continuously send queries, causing a constant workload, setting a baseline and buying a commitment provides constant capacity at a discounted price.

5. BigQuery autoscaling is subject to capacity availability. BigQuery attempts to meet customer capacity demand based on historical usage. To achieve capacity guarantees, you can set an optional slot baseline, which is the number of guaranteed slots in a reservation. With baselines, slots are immediately available and you pay for them whether you use them or not. To ensure capacity is available for large, inorganic demands, such as high-traffic holidays, contact the BigQuery team several weeks in advance.

6. Baseline slots are always charged. If a capacity commitment expires, you might need to manually adjust the amount of baseline slots in your reservations to avoid any unwanted charges. For example, consider that you have a 1-year commitment with 100 slots and a reservation with 100 baseline slots. The commitment expires and doesn't have a renewal plan. Once the commitment expires, you pay for 100 baseline slots at the pay as you go rate.

Monitor autoscaling

For information about monitoring slot usage and job performance with autoscaling, see Monitor autoscaling.

Excess slot usage

When a job holds onto slots for too long, it can receive an unfair share of slots. To prevent delays, BigQuery allows other jobs to borrow additional slots, resulting in periods of total slot use above your specified slot capacity. Any excess slot usage is attributed only to the jobs that receive more than their fair share.

The excess slots are not billed directly to you. Instead, jobs continue to run and accrue slot usage at their fair share until all of their excess usage is covered by your allocated capacity. Excess slots are excluded from reported slot usage with the exception of certain detailed execution statistics.

Note that some preemptive borrowing of slots can occur to reduce future delays and to provide other benefits such as reduced slot cost variability and reduced tail latency. Slot borrowing is limited to a small fraction of your total slot capacity.