Optimization and Best Practices

This topic explains some best practices of working with SQreamDB.

See also our Monitoring Query Performance guide for more information.

local:

Table design

Using DATE and DATETIME Data Types

When creating tables with dates or timestamps, using the purpose-built DATE and DATETIME types over integer types or TEXT will bring performance and storage footprint improvements, and in many cases huge performance improvements (as well as data integrity benefits). SQreamDB stores dates and datetimes very efficiently and can strongly optimize queries using these specific types.

Avoiding Data flattening and Denormalization

SQreamDB executes JOIN operations very effectively. It is almost always better to JOIN tables at query-time rather than flatten/denormalize your tables.

This will also reduce storage size and reduce row-lengths.

We highly suggest using INT or BIGINT as join keys, rather than a TEXT or STRING type.

Converting Foreign Tables to Native Tables

SQreamDB’s native storage is heavily optimized for analytic workloads. It is always faster for querying than other formats, even columnar ones such as Parquet. It also enables the use of additional metadata to help speed up queries, in some cases by many orders of magnitude.

You can improve the performance of all operations by converting Foreign Tables into native tables by using the CREATE TABLE AS syntax.

For example,

CREATE TABLE native_table AS SELECT * FROM foreign_table;

The one situation when this wouldn’t be as useful is when data will be only queried once.

Leveraging Column Data Information

Knowing the data types and their ranges can help design a better table.

Appropriately Using NULL and NOT NULL

For example, if a value cannot be missing (or NULL), specify a NOT NULL constraint on the columns.

Not only does specifying NOT NULL save on data storage, it lets the query compiler know that a column cannot have a NULL value, which can improve query performance.

Sorting

Data sorting is an important factor in minimizing storage size and improving query performance.

  • Minimizing storage saves on physical resources and increases performance by reducing overall disk I/O. Prioritize the sorting of low-cardinality columns. This reduces the number of chunks and extents that SQreamDB reads during query execution.

  • Where possible, sort columns with the lowest cardinality first. Avoid sorting TEXT columns with lengths exceeding 50 characters.

  • For longer-running queries that run on a regular basis, performance can be improved by sorting data based on the WHERE and GROUP BY parameters. Data can be sorted during insert by using external_tables or by using CREATE TABLE AS.

Query Best Practices

This section describes best practices for writing SQL queries.

Reducing Datasets Before Joining Tables

Reducing the input to a JOIN clause can increase performance. Some queries benefit from retrieving a reduced dataset as a subquery prior to a join.

For example,

SELECT store_name, SUM(amount)
FROM store_dim AS dim INNER JOIN store_fact AS fact ON dim.store_id=fact.store_id
WHERE p_date BETWEEN '2018-07-01' AND '2018-07-31'
GROUP BY 1;

Can be rewritten as

SELECT store_name, sum_amount
FROM store_dim AS dim INNER JOIN
   (SELECT SUM(amount) AS sum_amount, store_id
   FROM store_fact
   WHERE p_date BETWEEN '2018-07-01' AND '2018-07-31'
   group by 2) AS fact
ON dim.store_id=fact.store_id;

Using ANSI JOIN

SQreamDB prefers the ANSI JOIN syntax. In some cases, the ANSI JOIN performs better than the non-ANSI variety.

For example, this ANSI JOIN example will perform better:

ANSI JOIN will perform better
SELECT p.name, s.name, c.name
FROM  "Products" AS p
JOIN  "Sales" AS s
  ON  p.product_id = s.sale_id
JOIN  "Customers" as c
  ON  s.c_id = c.id AND c.id = 20301125;

This non-ANSI JOIN is supported, but not recommended:

Non-ANSI JOIN may not perform well
SELECT p.name, s.name, c.name
FROM "Products" AS p, "Sales" AS s, "Customers" as c
WHERE p.product_id = s.sale_id
  AND s.c_id = c.id
  AND c.id = 20301125;

Using High-Selectivity hint

Selectivity is the ratio of cardinality to the number of records of a chunk. We define selectivity as \(\frac{\text{Distinct values}}{\text{Total number of records in a chunk}}\)

SQreamDB has a hint function called HIGH_SELECTIVITY, which is a function you can wrap a condition in.

The hint signals to SQreamDB that the result of the condition will be very sparse, and that it should attempt to rechunk the results into fewer, fuller chunks.

Use the high selectivity hint when you expect a predicate to filter out most values. For example, when the data is dispersed over lots of chunks (meaning that the data is not well-clustered).

For example,

SELECT store_name, SUM(amount) FROM store_dim
WHERE HIGH_SELECTIVITY(p_date = '2018-07-01')
GROUP BY 1;

This hint tells the query compiler that the WHERE condition is expected to filter out more than 60% of values. It never affects the query results, but when used correctly can improve query performance.

Tip

The HIGH_SELECTIVITY() hint function can only be used as part of the WHERE clause. It can’t be used in equijoin conditions, cases, or in the select list.

Read more about identifying the scenarios for the high selectivity hint in our Monitoring query performance guide.

Avoiding Aggregation Overflow

When using an INT or smaller type, the SUM and COUNT operations return a value of the same type. To avoid overflow on large results, cast the column up to a larger type.

For example

SELECT store_name, SUM(amount :: BIGINT) FROM store_dim
GROUP BY 1;

Prefer COUNT(*) and COUNT to Non-nullable Columns

SQreamDB optimizes COUNT(*) queries very strongly. This also applies to COUNT(column_name) on non-nullable columns. Using COUNT(column_name) on a nullable column will operate quickly, but much slower than the previous variations.

Returning Only Required Columns

Returning only the columns you need to client programs can improve overall query performance. This also reduces the overall result set, which can improve performance in third-party tools.

SQreamDB is able to optimize out unneeded columns very strongly due to its columnar storage.

Reducing Recurring Compilation Time

Saved Queries are compiled when they are created. The query plan is saved in SQreamDB’s metadata for later re-use.

Saved query plans enable reduced compilation overhead, especially with very complex queries, such as queries with lots of values in an IN predicate.

When executed, the saved query plan is recalled and executed on the up-to-date data stored on disk.

Reducing JOIN Complexity

Filter and reduce table sizes prior to joining on them

SELECT store_name,
       SUM(amount)
FROM dimention dim
  JOIN fact ON dim.store_id = fact.store_id
WHERE p_date BETWEEN '2019-07-01' AND '2019-07-31'
GROUP BY store_name;

Can be rewritten as:

SELECT store_name,
       sum_amount
FROM dimention AS dim
  INNER JOIN (SELECT SUM(amount) AS sum_amount,
                     store_id
              FROM fact
              WHERE p_date BETWEEN '2019-07-01' AND '2019-07-31'
              GROUP BY store_id) AS fact ON dim.store_id = fact.store_id;

Data Loading Considerations

Using Natural Data Sorting

Very often, tabular data is already naturally ordered along a dimension such as a timestamp or area.

This natural order is a major factor for query performance later on, as data that is naturally sorted can be more easily compressed and analyzed with SQreamDB’s metadata collection.

For example, when data is sorted by timestamp, filtering on this timestamp is more effective than filtering on an unordered column.

Natural ordering can also be used for effective DELETE operations.

Use the Monitoring Query Performance guide to learn about built-in monitoring utilities. The guide also gives concrete examples for improving query performance.