Predicting database index hits

In a large application, it’s fairly easy to introduce code that produces poor or unoptimized SQL queries. As a developer, nothing stops you from writing this kind of a background job:

class MyJob
  def perform(params)
    products = Product.where(active: true, created_at: 1.month.ago..1.week.ago, user_id: params[:user_id])
    products.find_each do |product|
      # do something with the product
    end
  end
end

Only after the code goes through review and CI pipeline, when it’s merged and deployed, you’ll find out that the SQL query that the relation produces doesn’t hit any index - or hits one that is inefficient. The query would time out and the job would crash with an exception.

Some cases could be caught by peer code review but humans are not perfect in predicting query efficiency. This is better done by machines.

You could use EXPLAIN to see what indexes the query hits and what’s the cost of the query, but results of EXPLAIN depend of the actual dataset and may be completely different in production and local environments. That makes it impossible to check queries using EXPLAIN as a part of CI, without giving it access to the production dataset.

Let’s try a manual lookup of indexes for a given table by running SHOW INDEX FROM products. With a simple Ruby script we can compile a list of indexes and columns that they cover:

require 'bundler/setup'
require 'mysql2'

client = Mysql2::Client.new(host: "localhost", username: "root", database: "demo")
res = client.query "SHOW INDEX FROM products"

indexes = {}

res.each do |row|
  record = indexes[row["Key_name"]] ||= {}
  record[:columns] ||= []
  record[:columns][row["Seq_in_index"] - 1] = row["Column_name"]
end

puts indexes.inspect

Good news: ActiveRecord already provides an API to lookup indexes!

ActiveRecord::Base.connection.indexes(:product)

It looks like we could take columns from WHERE clause and match them with the columns covered by indexes. That way, we could build a dumb predictor of SQL query efficiency.

There’s a complication of ORDER BY that also affects a chosen index but we’ll simplify it for now.

If we somehow implemented it, we could reduce the rate of human errors and prevent developers from shipping code that won’t be able to efficiently run in production. In a large organization, that could save a few human hours per day.

Parsing queries

As we found, getting a list of indexes and columns that they cover is easy, especially with ActiveRecord. Now let’s see if we can identify columns mentioned in the WHERE clause.

With ActiveRecord, we could use where_values_hash.

Product.where(user_id: 42).where_values_hash
=> {"user_id"=>42}

But as we’ll learn later, it only returns values of exact matches and it doesn’t work for plain predicates and ranges:

Product.where("user_id IS NOT NULL").where(created_at: 1.month.ago..1.week.ago).where_values_hash
=> {}

If we look how where_values_hash is implemented, we’ll see that it reads Arel predicates. Let’s try hooking into Arel:

Product
  .where(active: true)
  .where("user_id IS NOT NULL")
  .where(created_at: 1.month.ago..1.week.ago)
  .where_clause
  .send(:predicates)
  .map(&:class)
=> [Arel::Nodes::Equality, String, Arel::Nodes::Between]

We could work with Arel::Nodes::Between and Arel::Nodes::Equality, but we’d still need to extract the column from "user_id IS NOT NULL" which is a string.

If we look broader, we’ll find something called libgda that has an AST parser of SQL queries. There’s even a Ruby binding for it. Let’s play with it:

class Visitor < GDA::Visitors::Visitor
  def visit_GDA_Nodes_Expr node
    puts "#{node.class}, #{node.value}, #{node.value.class}"
    super
  end
end

query = "SELECT id FROM products " \
  "WHERE user_id IS NOT NULL " \
  "AND active = 1 " \
  "AND created_at BETWEEN '2017-12-20 20:57:57' AND '2018-01-19 20:57:57'"

parser = GDA::SQL::Parser.new
result = parser.parse(query)
Visitor.new.accept(result.ast)

Which gives the following output:

GDA::Nodes::Expr, id, String
GDA::Nodes::Expr, products, String
GDA::Nodes::Expr, NULL, String
GDA::Nodes::Expr, NULL, String
GDA::Nodes::Expr, user_id, String
GDA::Nodes::Expr, NULL, String
GDA::Nodes::Expr, active, String
GDA::Nodes::Expr, 1, String
GDA::Nodes::Expr, NULL, String
GDA::Nodes::Expr, created_at, String
GDA::Nodes::Expr, '2017-12-20 20:57:57', String
GDA::Nodes::Expr, '2018-01-19 20:57:57', String

You can see AST nodes that GDA extracted from the query. There are columns and values, but all of them are of GDA::Nodes::Expr type. It gets tricky to separate what is a column and what is a value. Either I missed something about it, either GDA is too low level for our purpose.

Conclusion

To continue experiments, I’ll probably use Arel and manually parse user_id IS NOT NULL predicates. That may give me “good enough” results as I’ll be able to run it against a large codebase to see how many false positive it will identify.

Stay tuned to learn about results!

Comments

comments powered by Disqus