Firebolt is built to be the most efficient cloud data warehouse for data-intensive applications. The workloads that run on Firebolt come in many shapes and sizes, and we’ve built a system that can support a wide variety of them. The documentation on these pages will teach you how to get the most out of Firebolt for your workload. We explain system internals so you understand what Firebolt can do out of the box. More importantly, we teach you about how you can control Firebolt’s behaviour. We explain the levers that Firebolt provides to optimize your workload, and how you can use Firebolt’s telemetry to understand whether your changes make things faster.

​
How we think about performance

We’re building Firebolt for engineers that run mission-critical analytics workloads. This shapes how we build the system.
  • Fast by default. We want Firebolt to work well for you out of the box. Getting good performance for a complex workload should be easier than on your current platform. Even without tuning your workload, you should see great price/performance.
  • Control when you need it. Firebolt can power your mission-critical, low-latency analytics workloads. If you take a workload to production, you need a system that gives you control and predictability. Because of this, we give you as much control as possible. You can control data layout, caching, query plans, and more. We provide extensive observability so you can properly tune your workload.
  • Airbags included. Having control over a system is only great if you can test changes in a safe way. We have a modern architecture where metadata, storage, and compute are fully decoupled. This allows you to test changes to your workload in a fully isolated environment. When we expose levers, we give you a way to test them without impacting your production workloads.

​
The query lifecycle

This section describes the lifecycle of a SQL query in Firebolt.
1

Query Routing

When you send a query, it enters the Firebolt gateway. The gateway decides how to route the query within your Firebolt engine. If you use multiple clusters to scale for concurrency, the gateway chooses the cluster with the least load. It also decides which node in the cluster should be the coordinator for the query.
2

Query Parsing

Once the coordinator receives the query, it parses it. During parsing, we extract the databases that are required to execute the query.
3

Metadata Lookup

After resolving the database names, the coordinator contacts Firebolt’s metadata service to get the current metadata state. Using this lookup, Firebolt implements ACID transactions and strong consistency. There are a variety of caches on the coordinator to make this fast. We only retrieve the metadata delta since the last query.
4

Query Planning

Query planning converts the unoptimized SQL query into an optimized query plan that can be executed.
  • Validation. Once the metadata is retrieved, Firebolt validates the query. We check that it references valid objects, and that the user has the correct RBAC permissions.
  • Logical Planning. We then optimize the query plan. For this, we first apply 150+ rule-based optimizations to normalize the query. These apply optimizations such as pushing down filters and eliminating redundant joins. We then apply cost-based optimizations for things like aggregate placement and join ordering. Statistics for cost-based optimizations come from metadata and historical query telemetry.
  • Physical Planning. We then convert the logical plan into a physical plan. We insert the minimal number of shuffles to ensure that the query can scale out to all nodes in the cluster. We also select indices for data pruning, and apply advanced optimizations such as pruning across joins. Finally, the query is broken into independent stages. These stages are serialized and sent to the nodes in the cluster for execution.
5

Query Execution

The nodes in the cluster receive the stages of the query plan. They set up the computation primitives for Firebolt’s multi-threaded, vectorized execution engine. Data is passed between stages using an optimized shuffle primitive.
6

Storage Access

During query execution, we perform table scans. The query planner chose matching indices in step 4. These are now leveraged to improve scan performance. Scans use multi-tiered caching. Cold data is read from S3 and cached in SSD and main-memory.

​
How you can control Firebolt

There are different ways to control Firebolt. By knowing what types of levers we support, you get full control of Firebolt. Levers can be tied to the storage format, give you workload-level control, or control the behaviour of a specific query.

​
Schema-level control

Some levers are tied to your schema objects. These levers give you control over your physical data layout. Any engine and query that uses these schema objects will be affected. For example, you get fine-grained control over managed tables in Firebolt. 
CREATE FACT TABLE lineitem (
  l_orderkey BIGINT NOT NULL, 
  l_shipdate DATE NOT NULL,
  [...]
  l_comment TEXT NOT NULL) 
PARTITION BY EXTRACT(MONTH FROM l_shipdate)
PRIMARY INDEX l_orderkey;
You can control whether a table is sharded or replicated across the nodes in your cluster by declaring it as a FACT or DIMENSION table. In addition, you can define a PARTITION BY clause. Rows with different partition values go into different data files (tablets). The PRIMARY INDEX defines the sort order within each tablet. Firebolt can prune data using sparse indices on your partitioning and sorting keys.

​
Engine-level control

Some levers are tied to the engine object. These levers usually allow you to control workload-level optimizations. An example is the control we give you over AUTO VACUUM. When auto vacuum is turned on, an engine compacts data files (tablets) in the background. This leads to larger tablets which in turn leads to more efficient reads. 
ALTER ENGINE <engine_name> SET AUTO_VACUUM = OFF;
By turning auto vacuum on and off at the engine level, you can control how compute resources are allocated on that engine. You can either devote some resources to storage optimization, or make sure all resources go to user-sent queries.

​
Query-level control

Some levers give you control at a per-query basis. Look at our system settings for an exhaustive list of all settings in Firebolt. Compared to engine-level control, settings allow you to change the behaviour of a single query. Different queries running on the same engine can have different settings. For example, you can control the target tablet size for an INSERT query. This gives you fine-grained control over whether you want fast inserts that produce smaller tablets, or longer-running inserts that give you very large tablets. 
-- Option 1: Use an explicit SET command before running the query.
SET tablet_min_size_bytes = 4294967296;
SET tablet_max_size_bytes = 4294967296;
INSERT INTO playstats 
  SELECT * FROM READ_PARQUET(URL => 's3://firebolt-publishing-public/help_center_assets/firebolt_sample_dataset/playstats/*.parquet')

-- Option 2: Attach the settings directly to the query using WITH.
-- This is especially useful when you want to enumerate different setting combinations for benchmarking.
INSERT INTO playstats 
  SELECT * FROM READ_PARQUET(URL => 's3://firebolt-publishing-public/help_center_assets/firebolt_sample_dataset/playstats/*.parquet')
WITH (tablet_min_size_bytes = 4294967296, tablet_max_size_bytes = 4294967296);

​
Syntax-level control

Sometimes, you want to control individual fragments of your query. In these cases, Firebolt gives you syntax-level control. This is the most fine-grained way to control Firebolt’s behaviour. An example are MATERIALIZED CTEs. These give you control over how exactly a CTE is treated by our optimizer and runtime. 
-- By default CTEs are not materialized. This means they are not an optimization
-- boundary. 
-- This is useful if e.g. different predicates can be pushed into different
-- occurrences of the same CTE in the main query.
WITH regular_cte AS (
  <regular_cte_definition>
),
-- A materialized CTE is an optimization barrier. It's computed once and the result
-- is re-used within the query.
-- This is useful if e.g. the CTE performs a very expensive computation with a small
-- result set that's reused multiple times within the main query.
WITH materialized_cte AS MATERIALIZED (
  <materialized_cte_definition>
)
<main_query>