Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Adding page for CBO #24736

Open
wants to merge 10 commits into
base: master
Choose a base branch
from
Open
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
Original file line number Diff line number Diff line change
Expand Up @@ -14,9 +14,9 @@ type: docs

## Synopsis

ANALYZE collects statistics about the contents of tables in the database, and stores the results in the `pg_statistic` system catalog. These statistics help the query planner to determine the most efficient execution plans for queries.
ANALYZE collects statistics about the contents of tables in the database, and stores the results in the [pg_statistic](../../../../../architecture/system-catalog/#data-statistics), [pg_class](../../../../../architecture/system-catalog/#schema), and [pg_stat_all_tables](../../../../../architecture/system-catalog/#table-activity) system catalogs. These statistics help the query planner to determine the most efficient execution plans for queries.

The statistics are also used by the YugabyteDB [cost-based optimizer](../../../../../reference/configuration/yb-tserver/#yb-enable-base-scans-cost-model) (CBO) to create optimal execution plans for queries. When run on up-to-date statistics, CBO provides performance improvements and can reduce or eliminate the need to use hints or modify queries to optimize query execution.
The statistics are also used by the YugabyteDB [cost-based optimizer](../../../../../architecture/query-layer/planner-optimizer) (CBO) to create optimal execution plans for queries. When run on up-to-date statistics, CBO provides performance improvements and can reduce or eliminate the need to use hints or modify queries to optimize query execution.

{{< warning title="Run ANALYZE manually" >}}
Currently, YugabyteDB doesn't run a background job like PostgreSQL autovacuum to analyze the tables. To collect or update statistics, run the ANALYZE command manually.
Expand Down Expand Up @@ -51,6 +51,18 @@ Table name to be analyzed; may be schema-qualified. Optional. Omit to analyze al

List of columns to be analyzed. Optional. Omit to analyze all columns of the table.

## Resetting statistics

Over time, statistics can reach a point where they no longer represent the current workload accurately. Resetting allows you to measure the impact of recent changes, like optimizations or new queries, without the influence of historical data. Also when diagnosing issues, fresh statistics can help pinpoint current issues more effectively, rather than having to sift through historical data that may not be relevant.

The `yb_reset_analyze_statistics` function is a convenient helper that offers an easy way to clear statistics collected for a specific table or for all tables within a database. This function can be called as,

```sql
SELECT yb_reset_analyze_statistics ( table_oid );
```

If table_oid is NULL, this function resets the statistics for all the tables in the current database that the user can analyze.

## Examples

### Analyze a single table
Expand Down
46 changes: 36 additions & 10 deletions docs/content/preview/architecture/docdb/lsm-sst.md
Original file line number Diff line number Diff line change
Expand Up @@ -12,51 +12,77 @@ menu:
type: docs
---

A log-structured merge-tree (LSM tree) is a data structure and storage architecture used by [RocksDB](http://rocksdb.org/), the underlying key-value store of DocDB. LSM trees strike a balance between write and read performance, making them suitable for workloads that involve both frequent writes and efficient reads.
A [log-structured merge-tree (LSM tree)](https://en.wikipedia.org/wiki/Log-structured_merge-tree) is a data structure and storage architecture used by [RocksDB](http://rocksdb.org/), the underlying key-value store of DocDB. LSM trees strike a balance between write and read performance, making them suitable for workloads that involve both frequent writes and efficient reads.

The core idea behind an LSM tree is to separate the write and read paths, allowing writes to be sequential and buffered in memory making them faster than random writes, while reads can still access data efficiently through a hierarchical structure of sorted files on disk.

An LSM tree has 2 primary components - Memtable and SSTs. Let's look into each of them in detail and understand how they work during writes and reads.
An LSM tree has 2 primary components - [Memtable](#memtable) and [Sorted String Tables (SSTs)](#sst). Let's look into each of them in detail and understand how they work during writes and reads.

{{<note>}}
Typically in LSMs there is a third component - WAL (Write ahead log). DocDB uses the Raft logs for this purpose. For more details, see [Raft log vs LSM WAL](../performance/#raft-vs-rocksdb-wal-logs).
{{</note>}}

## Comparison to B-tree

Most traditional databases (for example, MySQL, PostgreSQL, Oracle) have a [B-tree](https://en.wikipedia.org/wiki/B-tree) based storage system. But YugabyteDB had to chose an LSM based storage to build a highly scalable database for of the following reasons.
Most traditional databases (for example, MySQL, PostgreSQL, Oracle) have a [B-tree](https://en.wikipedia.org/wiki/B-tree)-based storage system. But Yugabyte chose LSM-based storage to build a highly scalable database for the following reasons:

- Write operations (insert, update, delete) are more expensive in a B-tree. As it involves random writes and in place node splitting and rebalancing. In an LSM-based storage, data is added to the [memtable](#memtable) and written onto a [SST](#sst) file as a batch.
- Write operations (insert, update, delete) are more expensive in a B-tree, requiring random writes and in-place node splitting and rebalancing. In LSM-based storage, data is added to the [memtable](#memtable) and written onto a [SST](#sst) file as a batch.
- The append-only nature of LSM makes it more efficient for concurrent write operations.

## Memtable

All new write operations (inserts, updates, and deletes) are written as key-value pairs to an in-memory data structure called a Memtable, which is essentially a sorted map or tree. The key-value pairs are stored in sorted order based on the keys. When the Memtable reaches a certain size, it is made immutable, which means no new writes can be accepted into that Memtable.
All new write operations (inserts, updates, and deletes) are written as key-value pairs to an in-memory data structure called a memtable, which is essentially a sorted map or tree. The key-value pairs are stored in sorted order based on the keys. When the memtable reaches a certain size, it is made immutable, which means no new writes can be accepted into that memtable.

The immutable Memtable is then flushed to disk as an SST (Sorted String Table) file. This process involves writing the key-value pairs from the Memtable to disk in a sorted order, creating an SST file. DocDB maintains one active Memtable, and utmost one immutable Memtable at any point in time. This ensures that write operations can continue to be processed in the active Memtable, when the immutable memtable is being flushed to disk.
## Flush to SST

The immutable [memtable](#memtable) is then flushed to disk as an [SST (Sorted String Table)](#sst) file. This process involves writing the key-value pairs from the memtable to disk in a sorted order, creating an SST file. DocDB maintains one active memtable, and at most one immutable memtable at any point in time. This ensures that write operations can continue to be processed in the active memtable while the immutable memtable is being flushed to disk.

## SST

Each SST (Sorted String Table) file is an immutable, sorted file containing key-value pairs. The data is organized into data blocks, which are compressed using configurable compression algorithms (for example, Snappy, Zlib). Index blocks provide a mapping between key ranges and the corresponding data blocks, enabling efficient lookup of key-value pairs. Filter blocks containing bloom filters allow for quickly determining if a key might exist in an SST file or not, skipping entire files during lookups. The footer section of an SST file contains metadata about the file, such as the number of entries, compression algorithms used, and pointers to the index and filter blocks.
Each SST file is an immutable, sorted file containing key-value pairs. The data is organized into data blocks, which are compressed using configurable compression algorithms (for example, Snappy, Zlib). Index blocks provide a mapping between key ranges and the corresponding data blocks, enabling efficient lookup of key-value pairs. Filter blocks containing bloom filters allow for quickly determining if a key might exist in an SST file or not, skipping entire files during lookups. The footer section of an SST file contains metadata about the file, such as the number of entries, compression algorithms used, and pointers to the index and filter blocks.

Each SST file contains a bloom filter, which is a space-efficient data structure that helps quickly determine whether a key might exist in that file or not, avoiding unnecessary disk reads.

{{<note>}}
Most LSMs organize SSTS into multiple levels, where each level contains one or more SST files. But DocDB maintains files in only one level (level0).
Most LSMs organize SSTs into multiple levels, where each level contains one or more SST files. But DocDB maintains files in only one level (level0).
{{</note>}}

Three core low-level operations are used to iterate through the data in SST files.

### Seek

The _seek_ operation is used to locate a specific key or position in an SST file or memtable. When performing a seek, the system attempts to jump directly to the position of the specified key. If the exact key is not found, seek positions the iterator at the closest key that is greater than or equal to the specified key, enabling efficient range scans or prefix matching.

### Next

The _next_ operation moves the iterator to the following key in sorted order. It is typically used for sequential reads or scans, where a query iterates over multiple keys, such as retrieving a range of rows. After a seek, a sequence of next operations can scan through keys in ascending order.

### Previous

The _previous_ operation moves the iterator to the preceding key in sorted order. It is useful for reverse scans or for reading records in descending order. This is important for cases where backward traversal is required, such as reverse range queries. For example, after seeking to a key near the end of a range, previous can be used to iterate through keys in descending order, often needed in order-by-descending queries.

## Write path

When new data is written to the LSM system, it is first inserted into the active Memtable. As the Memtable fills up, it is made immutable and written to disk as an SST file. Each SST file is sorted by key and contains a series of key-value pairs organized into data blocks, along with index and filter blocks for efficient lookups.
When new data is written to the LSM system, it is first inserted into the active memtable. As the memtable fills up, it is made immutable and written to disk as an SST file. Each SST file is sorted by key and contains a series of key-value pairs organized into data blocks, along with index and filter blocks for efficient lookups.

## Read Path

To read a key, the LSM tree first checks the Memtable for the most recent value. If not found, it checks the SST files and finds the key or determines that it doesn't exist. During this process, LSM uses the index and filter blocks in the SST files to efficiently locate the relevant data blocks containing the key-value pairs.
To read a key, the LSM tree first checks the memtable for the most recent value. If not found, it checks the SST files and finds the key or determines that it doesn't exist. During this process, LSM uses the index and filter blocks in the SST files to efficiently locate the relevant data blocks containing the key-value pairs.

## Delete path

Rather than immediately removing the key from SSTs, the delete operation marks a key as deleted using a tombstone marker, indicating that the key should be ignored in future reads. The actual deletion happens during [compaction](#compaction), when tombstones are removed along with the data they mark as deleted.

## Compaction

As data accumulates in SSTs, a process called compaction merges and sorts the SST files with overlapping key ranges producing a new set of SST files. The merge process during compaction helps to organize and sort the data, maintaining a consistent on-disk format and reclaiming space from obsolete data versions.

The [YB-TServer](../../yb-tserver/) manages multiple compaction queues and enforces throttling to avoid compaction storms. Although full compactions can be scheduled, they can also be triggered manually. Full compactions are also triggered automatically if the system detects tombstones and obsolete keys affecting read performance.

{{<lead link="../../yb-tserver/">}}
To learn more about YB-TServer compaction operations, refer to [YB-TServer](../../yb-tserver/)
{{</lead>}}

## Learn more

- [Blog: Background Compactions in YugabyteDB](https://www.yugabyte.com/blog/background-data-compaction/#what-is-a-data-compaction)
24 changes: 8 additions & 16 deletions docs/content/preview/architecture/query-layer/_index.md
Original file line number Diff line number Diff line change
Expand Up @@ -59,32 +59,25 @@ Views are realized during this phase. Whenever a query against a view (that is,

### Planner

YugabyteDB needs to determine the most efficient way to execute a query and return the results. This process is handled by the query planner/optimizer component.
The YugabyteDB query planner plays a crucial role in efficiently executing SQL queries across multiple nodes. It extends the capabilities of the traditional single node query planner to handle distributed data and execution.

The planner first analyzes different ways a query can be executed based on the available data and indexes. It considers various strategies like scanning tables sequentially or using indexes to quickly locate specific data.

If the query involves joining multiple tables, the planner evaluates different techniques to combine the data:
After determining the optimal plan, the planner generates a detailed execution plan with all the necessary steps, such as scanning tables, joining data, filtering rows, sorting, and computing expressions.

- Nested loop join: Scanning one table for each row in the other table. This can be efficient if one table is small or has a good index.
- Merge join: Sorting both tables by the join columns and then merging them in parallel. This works well when the tables are already sorted or can be efficiently sorted.
- Hash join: Building a hash table from one table and then scanning the other table to find matches in the hash table.
For queries involving more than two tables, the planner considers different sequences of joining the tables to find the most efficient approach.
The execution plan is then passed to the query executor component, which carries out the plan and returns the final query results.

The planner estimates the cost of each possible execution plan and chooses the one expected to be the fastest, taking into account factors like table sizes, indexes, sorting requirements, and so on.

After the optimal plan is determined, YugabyteDB generates a detailed execution plan with all the necessary steps, such as scanning tables, joining data, filtering rows, sorting, and computing expressions. This execution plan is then passed to the query executor component, which carries out the plan and returns the final query results.

{{<note>}}
The execution plans are cached for prepared statements to avoid overheads associated with repeated parsing of statements.
{{</note>}}
{{<lead link="./planner-optimizer/">}}
To learn how the query planner decides the optimal path for query execution, see [Query Planner](./planner-optimizer/)
{{</lead>}}

### Executor

After the query planner determines the optimal execution plan, the query executor component runs the plan and retrieves the required data. The executor sends appropriate requests to the other YB-TServers that hold the needed data to performs sorts, joins, aggregations, and then evaluates qualifications and finally returns the derived rows.
After the query planner determines the optimal execution plan, the executor runs the plan and retrieves the required data. The executor sends requests to the other YB-TServers that hold the data needed to perform sorts, joins, and aggregations, then evaluates qualifications, and finally returns the derived rows.

The executor works in a step-by-step fashion, recursively processing the plan from top to bottom. Each node in the plan tree is responsible for fetching or computing rows of data as requested by its parent node.

For example, if the top node is a "Merge Join" node, it first requests rows from its two child nodes (the left and right inputs to be joined). The executor recursively calls the child nodes to get rows from them.
For example, if the top node is a "Merge Join" node, it first requests rows from its two child nodes (the left and right inputs to be joined). The executor recursively calls the child nodes to retrieve rows.

A child node may be a "Sort" node, which requests rows from its child, sorts them, and returns the sorted rows. The bottom-most child could be a "Sequential Scan" node that reads rows directly from a table.

Expand All @@ -93,4 +86,3 @@ As the executor requests rows from each node, that node fetches or computes the
This process continues recursively until the top node has received all the rows it needs to produce the final result. For a `SELECT` query, these final rows are sent to the client. For data modification queries like `INSERT`, `UPDATE`, or `DELETE`, the rows are used to make the requested changes in the database tables.

The executor is designed to efficiently pull rows through the pipeline defined by the plan tree, processing rows in batches where possible for better performance.

Original file line number Diff line number Diff line change
Expand Up @@ -8,10 +8,9 @@ aliases:
- /preview/explore/ysql-language-features/join-strategies/
menu:
preview:
name: Join strategies
identifier: joins-strategies-ysql
parent: architecture-query-layer
weight: 100
weight: 200
type: docs
---

Expand Down
Loading