This query used to rely on a JOIN, effectively producing the following
SQL:
SELECT users.*
FROM users
LEFT OUTER JOIN emails ON emails.user_id = users.id
WHERE (users.email = X OR emails.email = X)
LIMIT 1;
The use of a JOIN means having to scan over all Emails and users, join
them together and then filter out the rows that don't match the criteria
(though this step may be taken into account already when joining).
In the new setup this query instead uses a sub-query, producing the
following SQL:
SELECT *
FROM users
WHERE id IN (select user_id FROM emails WHERE email = X)
OR email = X
LIMIT 1;
This query has the benefit that it:
1. Doesn't have to JOIN any rows
2. Only has to operate on a relatively small set of rows from the
"emails" table.
Since most users will only have a handful of Emails associated
(certainly not hundreds or even thousands) the size of the set returned
by the sub-query is small enough that it should not become problematic.
Performance of the old versus new version can be measured using the
following benchmark:
# Save this in ./bench.rb
require 'benchmark/ips'
email = 'yorick@gitlab.com'
def User.find_by_any_email_old(email)
user_table = arel_table
email_table = Email.arel_table
query = user_table.
project(user_table[Arel.star]).
join(email_table, Arel::Nodes::OuterJoin).
on(user_table[:id].eq(email_table[:user_id])).
where(user_table[:email].eq(email).or(email_table[:email].eq(email)))
find_by_sql(query.to_sql).first
end
Benchmark.ips do |bench|
bench.report 'original' do
User.find_by_any_email_old(email)
end
bench.report 'optimized' do
User.find_by_any_email(email)
end
bench.compare!
end
Running this locally using "bundle exec rails r bench.rb" produces the
following output:
Calculating -------------------------------------
original 1.000 i/100ms
optimized 93.000 i/100ms
-------------------------------------------------
original 11.103 (± 0.0%) i/s - 56.000
optimized 948.713 (± 5.3%) i/s - 4.743k
Comparison:
optimized: 948.7 i/s
original: 11.1 i/s - 85.45x slower
In other words, the new setup is 85x faster compared to the old setup,
at least when running this benchmark locally.
For GitLab.com these improvements result in User.find_by_any_email
taking only ~170 ms to run, instead of around 800 ms. While this is
"only" an improvement of about 4.5 times (instead of 85x) it's still
significantly better than before.
Fixes#3242
This cuts down the time it takes to sort issues of a milestone by about
10x. In the previous setup the code would run a SQL query for every
issue that had to be sorted. The new setup instead runs a single SQL
query to update all the given issues at once.
The attached benchmark used to run at around 60 iterations per second,
using the new setup this hovers around 600 iterations per second. Timing
wise a request to update a milestone with 40-something issues would take
about 760 ms, in the new setup this only takes about 130 ms.
Fixes#3066
Improve User.by_login performance
This greatly speeds up the performance of `User.by_login`. I adopted some changes from @haynes in this patch, the credits go to him for coming up with those originally.
Fixes#2341
See merge request !1545
By comparing objects in Ruby we can greatly improve the performance of
this method. In the worst case (should no data be eager loaded) this
will run the same amount of queries as before, in the best case (when
data _is_ eager loadeD) it requires no queries at all.
The added benchmark used to produce around 273 iterations per second.
With this commit this has been increased to almost 40 000 iterations per
second: a speedup of roughly 145 times.
Combined with eager loading Note associations this results in about 30
queries less when viewing a single issue, this in turn cuts down the
loading time by 30-40%.
Performance is improved in two steps:
1. On PostgreSQL an expression index is used for checking lower(email)
and lower(username).
2. The check to determine if we're searching for a username or Email is
moved to Ruby. Thanks to @haynes for suggesting and writing the
initial implementation of this.
Moving the check to Ruby makes this method an additional 1.5 times
faster compared to doing the check in the SQL query.
With performance being improved I've now also tweaked the amount of
iterations required by the User.by_login benchmark. This method now runs
between 900 and 1000 iterations per second.
By using a JOIN we can remove the need for using 2 separate queries to
find a project by its namespace. Combined with an index (only needed for
PostgreSQL) this reduces the query time from ~245 ms (~520 ms for the
first call) down to roughly 10 ms (~15 ms for the first call).
This class method can be used in "describe" blocks to specify the
subject of a benchmark. This lets you write:
benchmark_subject { Foo }
instead of:
benchmark_subject { -> { Foo } }
This benchmark suite uses benchmark-ips
(https://github.com/evanphx/benchmark-ips) behind the scenes. Specs can
be turned into benchmark specs by setting "benchmark" to "true" in the
top-level describe block like so:
describe SomeClass, benchmark: true do
end
Writing benchmarks can be done using custom RSpec matchers, for example:
describe MaruTheCat, benchmark: true do
describe '#jump_in_box' do
it 'should run 1000 iterations per second' do
maru = described_class.new
expect { maru.jump_in_box }.to iterate_per_second(1000)
end
end
end
By default the "iterate_per_second" expectation requires a standard
deviation under 30% (this is just an arbitrary default for now). You can
change this by chaining "with_maximum_stddev" on the expectation:
expect { maru.jump_in_box }.to iterate_per_second(1000)
.with_maximum_stddev(10)
This will change the expectation to require a maximum deviation of 10%.
Alternatively you can use the it block style to write specs:
describe MaruTheCat, benchmark: true do
describe '#jump_in_box' do
subject { -> { described_class.new } }
it { is_expected.to iterate_per_second(1000) }
end
end
Because "iterate_per_second" operates on a block, opposed to a static
value, the "subject" method must return a Proc. This looks a bit goofy
but I have been unable to find a nice way around this.
