Conseils pour travailler avec des bases de données normalisées complexes

Nous avons tous appris les avantages de la normalisation de nos données. Je ne vais donc pas vous ennuyer avec ces détails, mais pour résumer :

La normalisation est le processus d'organisation des données dans une base de données. Cela comprend la création de tables et l'établissement de relations entre ces tables selon des règles conçues à la fois pour protéger les données et pour rendre la base de données plus flexible en éliminant la redondance et les dépendances incohérentes.

Microsoft 365 - Description de la normalisation

Pour être honnête, la normalisation ne m'a jamais vraiment traversé l'esprit jusqu'à récemment, lorsque j'ai dû gérer plusieurs applications existantes qui étaient « hautement normalisées ». Et quand je dis « hautement normalisé », je veux dire « TRÈS NORMALISÉ » – au point que cela n’a plus de sens. Ce qui m'a rappelé cet article étonnant de Coding Horror : Peut-être que la normalisation n'est pas normale.

Le problème est que, à moins d’être vraiment chanceux, vous n’aurez pas à vous soucier de choses comme celle-ci. Au lieu d’en parler de manière hypothétique, passons en revue un scénario spécifique et essayons différentes techniques pour comprendre les complexités de ce sujet. Une fois que nous avons examiné ce scénario, parlons des détails techniques pour mieux comprendre pourquoi les architectures hautement normalisées pourraient être problématiques et passons en revue les optimisations que nous pouvons envisager pour améliorer notre expérience.

? Vous pouvez consulter le code de cet article ici.

Vous travaillez sur un système de gestion des stocks existant et à grande échelle, basé sur SASS (software as a service). Le système se compose d'articles en stock et chaque article en stock a une catégorie, un fournisseur, un entrepôt et divers attributs. Un client a demandé un rapport et ce rapport doit afficher les détails de l'article, y compris le nom de son fournisseur et le nom de son entrepôt.

Voici un schéma simplifié, sans la multi-location (juste pour garder les choses simples) :

Chaque élément fait référence à des entrées dans les tableaux des catégories, des fournisseurs et des entrepôts. Les attributs de chaque élément sont stockés dans la table item_attributes. Tout cela a du sens et est assez facile à concocter :

CREATE TABLE items (
    id INT AUTO_INCREMENT PRIMARY KEY,
    name VARCHAR(255) NOT NULL,
    category_id INT,
    supplier_id INT,
    warehouse_id INT,
    created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
    updated_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP ON UPDATE CURRENT_TIMESTAMP,
    FOREIGN KEY (category_id) REFERENCES categories(id),
    FOREIGN KEY (supplier_id) REFERENCES suppliers(id),
    FOREIGN KEY (warehouse_id) REFERENCES warehouses(id)
);

CREATE TABLE categories (
    id INT AUTO_INCREMENT PRIMARY KEY,
    name VARCHAR(255) NOT NULL,
    created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
    updated_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP ON UPDATE CURRENT_TIMESTAMP
);

CREATE TABLE suppliers (
    id INT AUTO_INCREMENT PRIMARY KEY,
    name VARCHAR(255) NOT NULL,
    created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
    updated_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP ON UPDATE CURRENT_TIMESTAMP
);

CREATE TABLE warehouses (
    id INT AUTO_INCREMENT PRIMARY KEY,
    name VARCHAR(255) NOT NULL,
    location VARCHAR(255) NOT NULL,
    created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
    updated_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP ON UPDATE CURRENT_TIMESTAMP
);

CREATE TABLE item_attributes (
    id INT AUTO_INCREMENT PRIMARY KEY,
    item_id INT,
    attribute_name VARCHAR(255) NOT NULL,
    attribute_value VARCHAR(255) NOT NULL,
    created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
    updated_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP ON UPDATE CURRENT_TIMESTAMP,
    FOREIGN KEY (item_id) REFERENCES items(id)
);

-- To illustrate the denormalization strategy mentioned, here’s an example of a denormalized items_denormalized table:

CREATE TABLE items_denormalized (
    id INT AUTO_INCREMENT PRIMARY KEY,
    name VARCHAR(255) NOT NULL,
    category_name VARCHAR(255),
    supplier_name VARCHAR(255),
    warehouse_name VARCHAR(255),
    attribute_name VARCHAR(255),
    attribute_value VARCHAR(255),
    created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
    updated_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP ON UPDATE CURRENT_TIMESTAMP
);

CREATE INDEX idx_items_id ON items(id);
CREATE INDEX idx_categories_id ON categories(id);
CREATE INDEX idx_suppliers_id ON suppliers(id);
CREATE INDEX idx_warehouses_id ON warehouses(id);
CREATE INDEX idx_item_attributes_item_id ON item_attributes(item_id);

Données d'ensemencement

Pour tout travail de performance que nous effectuons, il est important de pouvoir reproduire l’échelle que nous prévoyons afin d’avoir une bonne idée des performances de notre application. C'est pourquoi j'ai élaboré le script d'amorçage suivant :

require 'faker'

def create_records(message, &block)
  puts "Creating #{message}."
  starting = Process.clock_gettime(Process::CLOCK_MONOTONIC)
  yield if block_given?
  ending = Process.clock_gettime(Process::CLOCK_MONOTONIC)
  elapsed = ending - starting
  puts "#{message.capitalize} created. #{elapsed}"
end

puts 'Truncating database...'
ActiveRecord::Tasks::DatabaseTasks.truncate_all
puts 'Database truncated.'

create_records('categories') do
  10.times do
    Category.create(name: Faker::Book.genre)
  end
end

create_records('suppliers') do
  25.times do
    Supplier.create(name: Faker::Company.name)
  end
end

create_records('warehouses') do
  1000.times do
    Warehouse.create(name: Faker::Company.name, location: Faker::Address.full_address)
  end
end

create_records('items') do
  categories = Category.all.to_a
  suppliers = Supplier.all.to_a
  warehouses = Warehouse.all.to_a

  items = 100_000.times.map do
    {
      name: Faker::Commerce.product_name,
      category_id: categories.sample.id,
      supplier_id: suppliers.sample.id,
      warehouse_id: warehouses.sample.id
    }
  end

  items.each_slice(1000) do |batch|
    Item.insert_all(batch)
  end
end

create_records('item attributes') do
  items = Item.all

  # We'll bump this up later to 1_000_000 in order to see
  # the perf issues come up.
  item_attributes = 100_000.times.map do
    {
      attribute_name: Faker::Lorem.word,
      attribute_value: Faker::Lorem.word,
      item_id: items.sample.id
    }
  end

  item_attributes.each_slice(1000) do |batch|
    ItemAttribute.insert_all(batch)
  end
end

create_records('denormalized items') do
  items_with_associations = Item.includes(:category, :supplier, :warehouse)

  denormalized_items_attributes = []

  items_with_associations.find_each(batch_size: 1000) do |item|
    denormalized_items_attributes << {
      name: item.name,
      item_id: item.id,
      category_name: item.category.name,
      category_id: item.category.id,
      supplier_name: item.supplier.name,
      supplier_id: item.supplier.id,
      warehouse_name: item.warehouse.name,
      warehouse_id: item.warehouse.id,
      created_at: DateTime.now,
      updated_at: DateTime.now
    }
  end

  denormalized_items_attributes.each_slice(1000) do |batch|
    ItemDenormalized.insert_all(batch)
  end
end

Ce script d'amorçage permet de créer des enregistrements pour toutes nos entités. Vous pouvez affiner le script pour créer plus ou moins d'enregistrements afin de tester l'architecture. C'est exactement ce que nous allons faire dans quelques instants.

Maintenant, n'oubliez pas que cela fonctionnera sur votre ordinateur local, nous ne testons donc pas les ressources de niveau production ici. Espérons que vous puissiez disposer d'un environnement de niveau production pour expérimenter différentes stratégies, mais le but ici n'est pas de reproduire la production - mais de bien comprendre les complexités du travail avec des architectures hautement normalisées.

Lorsque nous exécutons les graines, nous obtiendrons les journaux suivants :

bundle exec rails db:seed
Truncating database...
Database truncated.
Creating categories.
Categories created. 3.226257999893278
Creating suppliers.
Suppliers created. 0.1299410001374781
Creating warehouses.
Warehouses created. 4.184017000021413
Creating items.
Items created. 7.629256000043824
Creating item attributes.
Item attributes created. 59.715396999847144
Creating denormalized items.
Denormalized items created. 12.066422999836504

Très bien, commençons à exécuter quelques requêtes.

Des performances ? Quelle performance ?!

Alors disons que je veux tous les articles sauf ceux de McDermott-Casper, un fournisseur qui a fait faillite. De plus, je ne veux pas d’éléments auxquels sont associés les attributs enim et/ou modi :

On peut écrire une requête, avec ActiveRecord, assez facilement comme ceci :

excluded_suppliers =
  Supplier
    .select('id')
    .where(name: "McDermott-Casper")
    .to_sql

excluded_attributes =
  ItemAttribute
    .select(:item_id)
    .where(attribute_name: ['enim', 'modi'])
    .to_sql

Item
  .distinct
  .select('items.id, items.name, categories.name AS category_name, suppliers.name AS supplier_name, warehouses.name AS warehouse_name')
  .joins(:category, :supplier, :warehouse)
  .left_outer_joins(:item_attributes)
  .where("items.supplier_id NOT IN (#{excluded_suppliers})")
  .where("items.id NOT IN(#{excluded_attributes})")
  .to_a

Les conditions d'exclusion des éléments en fonction de notre scénario sont utilisées dans les conditions WHERE en tant que sous-requêtes intégrées, tandis que nous joignons les attributs de catégorie, de fournisseur, d'entrepôt et (jointure externe gauche) pour garantir que nous récupérons uniquement les éléments correspondants à notre condition. .

Très bien, testons ça :

bundle exec rails c
Loading development environment (Rails 7.1.3.4)
irb(main):001* excluded_suppliers =
irb(main):002>   Supplier
irb(main):003>     .select('id')
irb(main):004>     .where(name: "McDermott-Casper")
irb(main):005>     .to_sql
=> "SELECT \"suppliers\".\"id\" FROM \"suppliers\" WHERE \"suppliers\".\"name\" = 'McDermott-Casper'"
irb(main):006* excluded_attributes =
irb(main):007>   ItemAttribute
irb(main):008>     .select(:item_id)
irb(main):009>     .where(attribute_name: ['enim', 'modi'])
irb(main):010>     .to_sql
=> "SELECT \"item_attributes\".\"item_id\" FROM \"item_attributes\" WHERE \"item_attributes\".\"attribute_name\" IN ('enim', 'modi')"
irb(main):011> Item
irb(main):012>   .distinct
irb(main):013>   .select('items.id, items.name, categories.name AS category_name, suppliers.name AS supplier_name, warehouses.name AS warehouse_name')
irb(main):014>   .joins(:category, :supplier, :warehouse)
irb(main):015>   .left_outer_joins(:item_attributes)
irb(main):016>   .where("items.supplier_id NOT IN (#{excluded_suppliers})")
irb(main):017>   .where("items.id NOT IN(#{excluded_attributes})")
irb(main):018>   .to_a
  Item Load (535.5ms)  SELECT DISTINCT items.id, items.name, categories.name AS category_name, suppliers.name AS supplier_name, warehouses.name AS warehouse_name FROM "items" INNER JOIN "categories" ON "categories"."id" = "items"."category_id" INNER JOIN "suppliers" ON "suppliers"."id" = "items"."supplier_id" INNER JOIN "warehouses" ON "warehouses"."id" = "items"."warehouse_id" LEFT OUTER JOIN "item_attributes" ON "item_attributes"."item_id" = "items"."id" WHERE (items.supplier_id NOT IN (SELECT "suppliers"."id" FROM "suppliers" WHERE "suppliers"."name" = 'McDermott-Casper')) AND (items.id NOT IN(SELECT "item_attributes"."item_id" FROM "item_attributes" WHERE "item_attributes"."attribute_name" IN ('enim', 'modi')))
=>

Génial ! Nous en sommes à des récupérations inférieures à la seconde.

Très bien. Voyons ce qui se passe lorsque nous augmentons le nombre d'attributs du système à… disons un million. Nous pouvons le faire en exécutant le code suivant, extrait du script seed :

  items = Item.all

  # We'll bump this up later to 1_000_000 in order to see
  # the perf issues come up.
  item_attributes = 900_000.times.map do
    {
      attribute_name: Faker::Lorem.word,
      attribute_value: Faker::Lorem.word,
      item_id: items.sample.id
    }
  end

  item_attributes.each_slice(1000) do |batch|
    ItemAttribute.insert_all(batch)
  end

Gardez maintenant à l'esprit que ce qui précède contenait 1 187 enregistrements d'attributs d'objet correspondant à enim ou modi.

irb(main):001* excluded_suppliers =
irb(main):002>   Supplier
irb(main):003>     .select('id')
irb(main):004>     .where(name: "McDermott-Casper")
irb(main):005>     .to_sql
irb(main):006> 
=> "SELECT \"suppliers\".\"id\" FROM \"suppliers\" WHERE \"suppliers\".\"name\" = 'McDermott-Casper'"
irb(main):007* excluded_attributes =
irb(main):008>   ItemAttribute
irb(main):009>     .select(:item_id)
irb(main):010>     .where(attribute_name: ['enim', 'modi'])
irb(main):011>     .to_sql
irb(main):012> 
=> "SELECT \"item_attributes\".\"item_id\" FROM \"item_attributes\" WHERE \"item_attributes\".\"attribute_name\" IN ('enim', 'modi')"
irb(main):013> Item
irb(main):014>   .distinct
irb(main):015>   .select('items.id, items.name, categories.name AS category_name, suppliers.name AS supplier_name, warehouses.name AS warehouse_name')
irb(main):016>   .joins(:category, :supplier, :warehouse)
irb(main):017>   .left_outer_joins(:item_attributes)
irb(main):018>   .where("items.supplier_id NOT IN (#{excluded_suppliers})")
irb(main):019>   .where("items.id NOT IN(#{excluded_attributes})")
irb(main):020>   .to_a
irb(main):021> 
  Item Load (3002.4ms)  SELECT DISTINCT items.id,

Waouh ! D'accord. Nous en sommes maintenant à 3 secondes.

Le problème ne fera qu'empirer à mesure que de plus en plus d'éléments seront ajoutés au système au fil du temps et, en relation, item_attributes continuera à avoir un impact sur cette requête spécifique. Lorsque 900 000 attributs supplémentaires ont été ajoutés, le nombre d'enregistrements correspondant à enim ou modi a augmenté. En fait nous sommes passés de 1 187 à 12 154 enregistrements.

This kind of scale is completely normal and really shouldn’t be unexpected. As the number of attributes for items can increase significantly over time in an inventory management system for all sorts of reasons. Ok, so more records were added - of course performance would be impacted. What exactly is happening?

Is normalization really the issue here?

I’m going to remove the joins to categories and warehouses:

irb(main):029> Item
irb(main):030>   .distinct
irb(main):031>   .select('items.id, items.name, suppliers.name AS supplier_name')
irb(main):032>   .joins(:supplier)
irb(main):033>   .left_outer_joins(:item_attributes)
irb(main):034>   .where("items.supplier_id NOT IN (#{excluded_suppliers})")
irb(main):035>   .where("items.id NOT IN(#{excluded_attributes})")
irb(main):036>   .to_a
irb(main):037>
  Item Load (1938.4ms)  SELECT DISTINCT items.id, items.name, suppliers.name AS supplier_name FROM "items" INNER JOIN "suppliers" ON "suppliers"."id" = "items"."supplier_id" LEFT OUTER JOIN "item_attributes" ON "item_attributes"."item_id" = "items"."id" WHERE (items.supplier_id NOT IN (SELECT "suppliers"."id" FROM "suppliers" WHERE "suppliers"."name" = 'McDermott-Casper')) AND (items.id NOT IN(SELECT "item_attributes"."item_id" FROM "item_attributes" WHERE "item_attributes"."attribute_name" IN ('enim', 'modi')))
=>

Ok, so yeah, we get a ~30% improvement just removing the join. Let's run an explain on these and try to understand what's going on.

Unique  (cost=80266.89..84016.89 rows=250000 width=99)
  ->  Sort  (cost=80266.89..80891.89 rows=250000 width=99)
        Sort Key: items.id, items.name, categories.name, suppliers.name, warehouses.name
        ->  Hash Join  (cost=20105.00..44177.93 rows=250000 width=99)
              Hash Cond: (items.warehouse_id = warehouses.id)
              ->  Hash Join  (cost=20066.50..43480.40 rows=250000 width=89)
                    Hash Cond: (items.supplier_id = suppliers.id)
                    ->  Hash Join  (cost=20030.63..42785.86 rows=250000 width=78)
                          Hash Cond: (items.category_id = categories.id)
                          ->  Hash Right Join  (cost=19998.80..42094.91 rows=250000 width=54)
                                Hash Cond: (item_attributes.item_id = items.id)
                                ->  Seq Scan on item_attributes  (cost=0.00..19471.00 rows=1000000 width=8)
                                ->  Hash  (cost=19686.30..19686.30 rows=25000 width=54)
                                      ->  Seq Scan on items  (cost=16933.30..19686.30 rows=25000 width=54)
                                            Filter: ((NOT (hashed SubPlan 1)) AND (NOT (hashed SubPlan 2)))
                                            SubPlan 1
                                              ->  Seq Scan on suppliers suppliers_1  (cost=0.00..24.38 rows=1 width=8)
"                                                    Filter: ((name)::text = 'McDermott-Casper'::text)"
                                            SubPlan 2
                                              ->  Gather  (cost=1000.00..16878.93 rows=11996 width=8)
                                                    Workers Planned: 2
                                                    ->  Parallel Seq Scan on item_attributes item_attributes_1  (cost=0.00..14679.33 rows=4998 width=8)
"                                                          Filter: ((attribute_name)::text = ANY ('{enim,modi}'::text[]))"
                          ->  Hash  (cost=19.70..19.70 rows=970 width=40)
                                ->  Seq Scan on categories  (cost=0.00..19.70 rows=970 width=40)
                    ->  Hash  (cost=21.50..21.50 rows=1150 width=27)
                          ->  Seq Scan on suppliers  (cost=0.00..21.50 rows=1150 width=27)
              ->  Hash  (cost=26.00..26.00 rows=1000 width=26)
                    ->  Seq Scan on warehouses  (cost=0.00..26.00 rows=1000 width=26)

The plan above is telling us the output of each join is funneled into the next one:

(items <> warehouses) -> (items <> suppliers) -> (items <> categories)

Because of the multiple joins, we essentially increase the performance impact as more data is spread out across your database, e.g. normalization.

Now, let’s look at the plan after we remove the joins:

Unique  (cost=73750.91..76250.91 rows=250000 width=49)
  ->  Sort  (cost=73750.91..74375.91 rows=250000 width=49)
        Sort Key: items.id, items.name, suppliers.name
        ->  Hash Join  (cost=20034.68..42789.45 rows=250000 width=49)
              Hash Cond: (items.supplier_id = suppliers.id)
              ->  Hash Right Join  (cost=19998.80..42094.91 rows=250000 width=38)
                    Hash Cond: (item_attributes.item_id = items.id)
                    ->  Seq Scan on item_attributes  (cost=0.00..19471.00 rows=1000000 width=8)
                    ->  Hash  (cost=19686.30..19686.30 rows=25000 width=38)
                          ->  Seq Scan on items  (cost=16933.30..19686.30 rows=25000 width=38)
                                Filter: ((NOT (hashed SubPlan 1)) AND (NOT (hashed SubPlan 2)))
                                SubPlan 1
                                  ->  Seq Scan on suppliers suppliers_1  (cost=0.00..24.38 rows=1 width=8)
"                                        Filter: ((name)::text = 'McDermott-Casper'::text)"
                                SubPlan 2
                                  ->  Gather  (cost=1000.00..16878.93 rows=11996 width=8)
                                        Workers Planned: 2
                                        ->  Parallel Seq Scan on item_attributes item_attributes_1  (cost=0.00..14679.33 rows=4998 width=8)
"                                              Filter: ((attribute_name)::text = ANY ('{enim,modi}'::text[]))"
              ->  Hash  (cost=21.50..21.50 rows=1150 width=27)
                    ->  Seq Scan on suppliers  (cost=0.00..21.50 rows=1150 width=27)

Ok, so we get a better query plan. Less joins, less data to scan and therefore more performance. However, doing this won't meet the requirements. Remember, the report needs the names of the associated suppliers and warehouses. Let's see what happens when we denormalize the data and simplify the lookup process.

irb(main):074* excluded_suppliers =
irb(main):075>   Supplier
irb(main):076>     .select('id')
irb(main):077>     .where(name: "McDermott-Casper")
irb(main):078>     .to_sql
irb(main):079> 
irb(main):080* excluded_attributes =
irb(main):081>   ItemAttribute
irb(main):082>     .select(:item_id)
irb(main):083>     .where(attribute_name: ['enim', 'modi'])
irb(main):084>     .to_sql
irb(main):085> 
irb(main):086> ItemDenormalized
irb(main):087>   .distinct
irb(main):088>   .select('items_denormalized.id as id, items_denormalized.category_name as category_name, items_denormalized.supplier_name as supplier_name, items_denormalized.warehouse_name as warehouse_name')
irb(main):089>   .joins(:supplier)
irb(main):090>   .left_outer_joins(:item_attributes)
irb(main):091>   .where("items_denormalized.supplier_id NOT IN (#{excluded_suppliers})")
irb(main):092>   .where("items_denormalized.item_id NOT IN(#{excluded_attributes})")
irb(main):093>   .to_a
irb(main):094> 
  ItemDenormalized Load (1107.3ms)  SELECT DISTINCT items_denormalized.id as id,

In this example, the lookup on the denormalized table performed similarly to when we removed the joins (1107.3ms v. 1938.4ms). The difference is that we have the category and warehouse names. Denormalization does introduce multiple complexities that need to be handled; such as redundancy and integrity of the data, e.g. what happens when categories are updated? or when warehouses are deleted?

Putting that aside though, we see that denormalization handles certain scenarios well when it comes to performance. We should consider it's benefits when building applications that will inevitably need to scale. In our example above, we can see with just a million records, we start to run into some performance bottlenecks.

Performance Bottlenecks

Let's think through what bottlenecks start to come into play after running through the examples above.

Complex Queries

Highly normalized schemas often require complex queries with multiple joins, which can be slow and resource-intensive.

SELECT DISTINCT
    items.id,
    items.name,
    categories.name AS category_name,
    suppliers.name AS supplier_name,
    warehouses.name AS warehouse_name
FROM
    "items"
    INNER JOIN "categories" ON "categories"."id" = "items"."category_id"
    INNER JOIN "suppliers" ON "suppliers"."id" = "items"."supplier_id"
    INNER JOIN "warehouses" ON "warehouses"."id" = "items"."warehouse_id"
    LEFT OUTER JOIN "item_attributes" ON "item_attributes"."item_id" = "items"."id"
WHERE (items.supplier_id NOT IN(
        SELECT
            "suppliers"."id" FROM "suppliers"
        WHERE
            "suppliers"."name" = 'McDermott-Casper'))
    AND(items.id NOT IN(
            SELECT
                "item_attributes"."item_id" FROM "item_attributes"
            WHERE
                "item_attributes"."attribute_name" IN('enim', 'modi')));

I wouldn't consider the above too complex, however, the conditions that execute subqueries can start to get complex when joining on joins. This happens a lot in large scale applications that have evolved over time. Again, normalization is great in an ideal world - but it is also important to understand what other complexities it introduces.

Increased I/O Operations

Each table lookup can lead to additional I/O operations, slowing down the overall query performance. When we start to talk through IO operations in the database, it's important to know, high level, why this is an important part of the puzzle. So let's dive into some issues that come up at scale.

Read/Write: Each join that involves disk-based temporary tables or large data sets will increase the number of disk reads and writes. This can cause a significant I/O load, especially in applications where the behavior is quite active (jobs, high traffic, etc.).

Buffer Pool Pressure: Joins can put pressure on the MySQL buffer pool, especially with larger data sets. When the buffer pool is full, MySQL has to evict pages to make room for new data, causing additional disk I/O.

Temporary Tables:  MySQL may create temporary tables to hold intermediate results during complex join operations. These temporary tables can be stored in memory or on disk, depending on their size. Disk-based temporary tables increase I/O operations, leading to slower performance.

Lock Contention

In a highly concurrent environment, frequent access and updates across multiple tables can lead to lock contention and further degrade performance.

Multiple Joins

Lock Types: MySQL uses different types of locks (e.g., shared, exclusive) depending on the operation. Complex queries with multiple joins can require various locks, leading to contention if different parts of the query need the same resources.

Row-Level vs. Table-Level Locks: InnoDB uses row-level locking, which is generally more efficient than table-level locking used by MyISAM. However, even row-level locks can cause contention if multiple transactions try to modify the same rows simultaneously.

Joins on Joins

Increased Lock Duration: Queries involving joins on joins often take longer to execute. The longer a transaction holds locks, the higher the chance of contention with other transactions.

Lock Escalation: Although InnoDB uses row-level locking, high contention can sometimes cause lock escalation, where the database engine escalates to table-level locks to manage the contention, leading to broader performance issues. This is typically due to non-existent and/or lacking indexes.

Lock Waits and Deadlocks

Lock Waits: When a transaction needs a lock held by another transaction, it must wait, leading to increased query execution time and potential timeouts.

Deadlocks: Complex queries with multiple joins increase the risk of deadlocks, where two or more transactions are waiting for each other’s locks, causing the database to automatically roll back one of the transactions to resolve the deadlock, typically the "victim" is rolled back.

Strategies for Optimization

To mitigate performance issues in highly normalized architectures, consider the following strategies:

Denormalization

The process for denormalizing data involves adding redundant data to tables to reduce the number of joins required. While this increases storage requirements and the risk of data anomalies, it can significantly improve read performance.

SELECT i.id, i.name, i.category_name, i.supplier_name, i.warehouse_name, i.attribute_value
FROM items_denormalized i
WHERE i.id = ?

In this example, the items_denormalized table combines data from the categories, suppliers, warehouses, and item_attributes tables, eliminating the need for multiple joins.

Indexing

Proper indexing can dramatically improve query performance. Ensure that all columns used in joins and WHERE clauses are indexed. Remember, an index is super important to prevent full table locks. Keep in mind, that even this will not help if temporary tables are created with your joins, which will NOT have indexes.

CREATE INDEX idx_items_id ON items(id);
CREATE INDEX idx_categories_id ON categories(id);
CREATE INDEX idx_suppliers_id ON suppliers(id);
CREATE INDEX idx_warehouses_id ON warehouses(id);
CREATE INDEX idx_item_attributes_item_id ON item_attributes(item_id);

Caching

Implement caching mechanisms to store frequently accessed data in memory, reducing the need for repeated database queries. There are multiple strategies for implementing caching, which will be covered in a different post, but these strategies can range from utilizing summary tables, to integrating different technologies that can store results temporarily.

# Example using Ruby on Rails with Redis cache
item = Rails.cache.fetch("item_#{id}", expires_in: 12.hours) do
  Item.includes(:category, :supplier, :warehouse, :item_attributes).find(id)
end

Query Optimization

Analyze and optimize your queries to ensure they are as efficient as possible. Use tools like MySQL’s EXPLAIN ANALYZE statement to understand the execution plan and identify bottlenecks.

EXPLAIN SELECT i.id, i.name, c.name AS category, s.name AS supplier, w.name AS warehouse, ia.attribute_value
FROM items i
JOIN categories c ON i.category_id = c.id
JOIN suppliers s ON i.supplier_id = s.id
JOIN warehouses w ON i.warehouse_id = w.id
JOIN item_attributes ia ON i.id = ia.item_id
WHERE i.id = 1;

Conclusion

Normalization is a powerful technique for maintaining data integrity, but it can lead to performance challenges in large-scale applications. Knowing the tradeoffs here can help you scale your application in the long term, considering denormalization as just another strategy to help scale. If denormalization is not favorable; consider reviewing indices (including composites), result caching and query optimization to improve performance. Thank you for reading and please reach out if you have any questions!

References

Microsoft 365 - Description of the database normalization basics

Coding Horror - Maybe Normalizing Isn't Normal

informIT - When You Can't Change a SQL Database Design

PureStorage - Denormalized vs. Normalized Data.

MySQL - Buffer Pool

MySQL - InnoDB Disk I/O

MySQL - Internal Temporary Table Use in MySQL

MySQL - Locks Set by Different SQL Statements in InnoDB

Percona - Understanding Hash Joins in MySQL 8

Percona - Horizontal Scaling in MySQL – Sharding Followup

PlanetScale - How to Scale your Database and when to Shard MySQL

Awesome - Database Design

Awesome - MySQL

Ce qui précède est le contenu détaillé de. pour plus d'informations, suivez d'autres articles connexes sur le site Web de PHP en chinois!