An unfair advantage: multi-tenant queues in Postgres

An interesting post on allocating jobs in queues “fair” for multi-tenants. The algorithm described does the job distribution at write time, leaving a simpler dequeueing process.

Ruby Heredocs

Good reference on how to use ruby HEREDOCs. Depending on how often you use them, it’s easy to forget the specific syntax.

When Do We Stop Finding New Music? A Statistical Analysis

Not a lot of complicated stats here, except some slicing and averaging. However, the conclusion resonates. Your musical tastes stop evolving. I certainly find many new less artists than I did before. That said, my experience is not exactly like what is described: While I still like the music that I listened to when I was 13-15, I listen more to music that I didn’t really get into until much later, in my late 20s and early 30s.

Postgres has support for creating views that are themselves updatable, as if they are tables. By default, this is only possible for simple views, as defined in the documentation. More advanced views, can also be updatable, with a bit more work. Lets examine how.

For the sake of this example¹, lets say we have a user_last_logins table that captures the last date and time that the user has logged into a system:

CREATE TABLE user_last_logins (
  user_id integer NOT NULL,
  logged_in_at timestamp NOT NULL
);
-- CREATE TABLE
-- Time: 10.102 ms

CREATE UNIQUE INDEX idex_user_id_on_user_last_logins
ON user_last_logins(user_id);
-- CREATE INDEX
-- Time: 7.783 ms

Somewhere in the application code, when the user logs in, we update the time that a user has logged in.

INSERT INTO user_last_logins(user_id, logged_in_at)
VALUES
  (1, NOW()), (2, NOW()), (3, NOW());
-- INSERT 0 3
-- Time: 5.007 ms

SELECT * FROM user_last_logins;
--  user_id |        logged_in_at
-- ---------+----------------------------
--        1 | 2024-04-29 14:36:11.765157
--        2 | 2024-04-29 14:36:11.765157
--        3 | 2024-04-29 14:36:11.765157
-- (3 rows)
--
-- Time: 0.481 ms


UPDATE user_last_logins
SET logged_in_at = NOW()
WHERE user_id = 1;
-- UPDATE 1
-- Time: 5.463 ms

SELECT * FROM user_last_logins;
--  user_id |        logged_in_at
-- ---------+----------------------------
--        2 | 2024-04-29 14:36:11.765157
--        3 | 2024-04-29 14:36:11.765157
--        1 | 2024-04-29 14:36:22.268137
-- (3 rows)
--
-- Time: 0.973 ms

Notice how the timestamp for the row with user_id = 1 was updated.

Now lets imagine that our requirements change. We are now set to record the time of every login by users, as opposed to the last time. We also have different client applications that might not be on the same schedule, and would like to keep the current database functionality intact. Our first order of business is to create our new table and populated with the information we have at hand:

BEGIN;

CREATE TABLE user_logins (
  user_id integer NOT NULL,
  logged_in_at timestamp NOT NULL
);

INSERT INTO user_logins (user_id, logged_in_at)
SELECT * FROM user_last_logins;

COMMIT;
-- BEGIN
-- Time: 0.129 ms
-- CREATE TABLE
-- Time: 4.720 ms
-- INSERT 0 3
-- Time: 0.842 ms
-- COMMIT
-- Time: 0.339 ms

SELECT * from user_logins;
--
--  user_id |        logged_in_at
-- ---------+----------------------------
--        2 | 2024-04-29 14:36:11.765157
--        3 | 2024-04-29 14:36:11.765157
--        1 | 2024-04-29 14:36:22.268137
-- (3 rows)
--
-- Time: 0.816 ms

The new table has the same structure as the old table, but has no unique index on user_id, because we want to allow multiple rows for each user. I am aware that I could have renamed the table instead of creating a new one and copying data. I choose to do it this way for didactical purposes.

It’s now possible to insert new records for each users:

INSERT INTO user_logins VALUES (1, NOW());
-- INSERT 0 1
-- Time: 1.613 ms

SELECT * from user_logins;
--  user_id |        logged_in_at
-- ---------+----------------------------
--        2 | 2024-04-29 14:36:11.765157
--        3 | 2024-04-29 14:36:11.765157
--        1 | 2024-04-29 14:36:22.268137
--        1 | 2024-04-29 14:37:23.696786
-- (4 rows)
--
-- Time: 0.505 ms

We still have a problem: There are clients to this application that can’t change their code right away. Let’s tackle first the clients that want to read an up-to-date user_last_logins relation:

BEGIN;

DROP TABLE user_last_logins;

CREATE OR REPLACE VIEW user_last_logins AS
SELECT user_id, MAX(logged_in_at) as logged_in_at
FROM user_logins
GROUP BY 1;

COMMIT;
-- BEGIN
-- Time: 0.103 ms
-- DROP TABLE
-- Time: 3.219 ms
-- CREATE VIEW
-- Time: 6.703 ms
-- COMMIT
-- Time: 1.338 ms

SELECT * FROM user_last_logins;
--  user_id |        logged_in_at
-- ---------+----------------------------
--        3 | 2024-04-29 14:36:11.765157
--        2 | 2024-04-29 14:36:11.765157
--        1 | 2024-04-29 14:37:23.696786
-- (3 rows)
--
-- Time: 1.273 ms

We’ve created a view that produced the same information that used to be in user_last_logins from the underlying user_logins table, with the same guarantees that a user_id will only show up in a single row. Read clients can continue working without a hitch. However, write clients won’t be able to update as before:

UPDATE user_last_logins
SET logged_in_at = NOW()
WHERE user_id = 1;
-- ERROR:  cannot update view "user_last_logins"
-- DETAIL:  Views containing GROUP BY are not automatically updatable.
-- HINT:  To enable updating the view, provide an INSTEAD OF UPDATE trigger or an unconditional ON UPDATE DO INSTEAD rule.
-- Time: 0.975 ms

Conceptually, we know that the underlying table has the same structure, and that the insert to the view can be forwarded to the underlying table. The error even gives us a hint at what to do: Use and INSTEAD OF UPDATE trigger.

BEGIN;

CREATE OR REPLACE FUNCTION instead_function_insert_user_last_logins() RETURNS TRIGGER AS
$BODY$
BEGIN

INSERT INTO user_logins(user_id, logged_in_at)
VALUES (NEW.user_id, NEW.logged_in_at);

RETURN NEW;
END;
$BODY$
LANGUAGE plpgsql;

CREATE TRIGGER instead_trigger_user_last_logins
INSTEAD OF UPDATE ON user_last_logins
FOR EACH ROW
EXECUTE PROCEDURE instead_function_insert_user_last_logins();

COMMIT;
-- BEGIN
-- Time: 0.077 ms
-- CREATE FUNCTION
-- Time: 6.508 ms
-- CREATE TRIGGER
-- Time: 2.049 ms
-- COMMIT
-- Time: 0.427 ms

We’ve defined a trigger that runs on each attempted update of user_last_logins and instead runs a function. The body of the function plainly inserts instead into user_logins. And because user_logins powers the user_last_logins view, the insert appears to work as requested.

UPDATE user_last_logins
SET logged_in_at = NOW()
WHERE user_id = 1;
-- UPDATE 1
-- Time: 6.153 ms

SELECT * FROM user_last_logins;
--  user_id |        logged_in_at
-- ---------+----------------------------
--        3 | 2024-04-29 14:36:11.765157
--        2 | 2024-04-29 14:36:11.765157
--        1 | 2024-04-29 14:38:39.369368
-- (3 rows)
--
-- Time: 0.696 ms


SELECT * FROM user_logins;
--  user_id |        logged_in_at
-- ---------+----------------------------
--        2 | 2024-04-29 14:36:11.765157
--        3 | 2024-04-29 14:36:11.765157
--        1 | 2024-04-29 14:36:22.268137
--        1 | 2024-04-29 14:37:23.696786
--        1 | 2024-04-29 14:38:39.369368
-- (5 rows)
--
-- Time: 0.826 ms

Now our existing clients can migrate to use user_logins at their own pace. For this particular case, we only dealt with INSTEAD OF UPDATE, but in the same vein we could have defined an INSTEAD OF INSERT trigger.

Conclusions

The INSTEAD OF trigger allows us a complex view to remain updatable. In this made-up-case this allows views to function as an abstraction for tables that no longer exist, and de-couples the need from client code changing at the same time as the database schema.

Bash Debugging

Another gem by Julia Evans. I always learn something from her zines, and this is no exception: You can have a step debugger in bash!

How Figma’s Databases Team Lived to Tell the Scale

Very interesting article on how Figma managed to scale Postgres. Having a DB proxy that checks the queries and routes to the correct shard (and even aggregates among shards) is wild.

The use of logical partitioning to prove their work before doing actual physical partitioning is very clever.

Before going out and building a custom replication scheme, remember that there are out-of-the-box solutions out there that most organizations are better choosing over custom solutions.

jnv: interactive JSON filter using jq

jnv looks like a nice tool for interactively exploring the a JSON file.

Testing ActiveRecord Transactions the Right Way

Very handy recipes for testing ActiveRecord transactions.

Useful Uses of cat

Using cat to start a pipelines is about composing commands: It makes it easier to build pipelines in steps. Technically, you could be adding an extra process that you don’t need, but in day-to-day unix pipe operations, the performance is never an issue.

Data Looks Better Naked

Nice visual animation of how removing stuff improves the design. The pie chart in particular was great!

In a previous post – On Class Structure – we discussed how to organize Ruby classes so that the logic is not buried in private methods, and instead is clear to readers what the class is responsible for doing. In this post, I want to expand on that a bit more and introduce result objects.

Let’s resume with our example:

class FindScore
  DEFAULT_SCORE = 0
  URL = 'http://example.com'.freeze

  def initialize(user, http_client = HTTParty)
    @user = user
    @http_client = http_client
  end

  def call
    make_api_request(@user, @http_client)
      .then { parse_response(_1) }
      .then { extract_score(_1) }
  end

  private

  def make_api_request(user, http_client = @http_client, url = URL)
    http_client.post(
      url,
      body: { user: user.id }
    )
  end

  def parse_response(response)
    JSON.parse(response.body)
  end

  def extract_score(response_body, default = DEFAULT_SCORE)
    response_body.fetch("score", default)
  end
end

That class doesn’t handle any errors. Each of the private methods can fail in different ways. For the sake of examples, lets say that we can encounter HTTP errors in make_api_request, the response may fail to be valid JSON or the response might have a different JSON shape than what we expect. One way to handle them is via exceptions or checking for specific conditions, and then ensuring that the value passed along is what the next step in our pipeline expects:

class FindScore
  DEFAULT_SCORE = 0
  URL = 'http://example.com'.freeze

  def initialize(user, http_client = HTTParty)
    @user = user
    @http_client = http_client
  end

  def call
    make_api_request(@user, @http_client)
      .then { parse_response(_1) }
      .then { extract_score(_1) }
  end

  private

  def make_api_request(user, http_client = @http_client, url = URL)
    response = http_client.post(
      url,
      body: { user: user.id }
    )

    response.ok? ? response.body : "{}"
  end

  def parse_response(response)
    JSON.parse(response.body)
  rescue JSON::ParserError
    {}
  end

  def extract_score(response_body, default = DEFAULT_SCORE)
    response_body.fetch("score", default)
  end
end

In that version, #make_api_request checks for a correct response, passing the response body to #parse_response. If the response is not successful however, it returns "{}", which is JSON that will be parsable by that response. In a similar manner, parsing JSON might raise JSON::ParseError. #parse_response rescues the exception, and returns a hash, as expected by #extract_score.

The code is now more resilient: It can handle some errors and recover from them by returning a value that can be used in the next method. However, these errors are being swallowed. What if we wanted to add some logging or metrics for each error, so we can understand our system better? One way, is to add a logging statement on the error branch of each method. I prefer another way, using result objects.

For our purposes a result object can either be a success or an error. In either case, it wraps another value, and it has some methods that act differently in each case. This object is known as a result monad, but let’s now dwell on that. Our result object will make it easier to write pipelines of method calls, without sacrificing error handling.

A very minimal implementation looks like this:

class Ok
  def initialize(value)
    @value = value
  end

  def and_then
    yield @value
  end

  def value_or(_other)
    @value
  end
end

class Error
  def initialize(error)
    @error = error
  end

  def and_then
    self
  end

  def value_or
    yield @error
  end
end

The polymorphic interface for Ok and Error has two methods: #and_then which is used to pipeline operations, and #value_or which is used to unwrap the value. Let’s see some examples:

Ok.new(1)
  .and_then { |n| Ok.new(n * 2) } # => 1 * 2 = 2
  .and_then { |n| Ok.new(n + 1) } # => 2 + 1 = 3
  .value_or(:error)
# => 3

Ok.new(1)
  .and_then { |n| Ok.new(n * 2) } # => 1 * 2 = 3
  .and_then { |n| Error.new("something went wrong") }
  .and_then { |n| Ok.new(n + 1) } # => Never called
  .and_then { |n| raise "Hell" } # => Never called either
  .value_or { :error }
# => :error

A chain of #and_then calls continue much like #then does, expecting a result object as a return value. However, if the return value at any point is an Error, subsequent blocks will not execute, and instead will continue returning the same result object. We then have a powerful way of constructing pipelines. Error handling can be left to the end.

Our class with error handling, can now be written as:

class FindScore
  DEFAULT_SCORE = 0
  URL = 'http://example.com'.freeze

  def initialize(user, http_client = HTTParty)
    @user = user
    @http_client = http_client
  end

  def call
    make_api_request(@user, @http_client)
      .and_then { parse_response(_1) }
      .and_then { extract_score(_1) }
      .value_or { |error_message|
        log.error "FindScore failed for #{@user}: #{error_message}"
        DEFAULT_SCORE
      }
  end

  private

  def make_api_request(user, http_client = @http_client, url = URL)
    response = http_client.post(
      url,
      body: { user: user.id }
    )

    response.ok? ? Ok.new(response.body) : Error.new("HTTP Status Code: #{response.status_code}")
  end

  def parse_response(body)
    Ok.new(JSON.parse(body))
  rescue JSON::ParserError => ex
    Error.new(ex.to_s)
  end

  def extract_score(parsed_json)
    score = parsed_json["score"]

    score.present? ? Ok.new(score) : Error.new("Score not found in response")
  end
end

Now, each method is responsible for returning either an Ok or and Error. The #call method is responsible for constructing the overall pipeline and handling the failure (i.e. returning a DEFAULT_SCORE), and with a single line, it also logs all errors.

This technique is quite powerful. The result objects are not limited to private class methods. Public methods can return them just as well. The Ok and Error implementation is quite minimal as a demonstration for this post. There are full-featured libraries out there (e.g. dry-rb), or you can roll your own pretty easily and expand the API to suit your needs (e.g. #ok?, #error?, #value!, #error, #fmap).

As I concluded in my previous post, writing Ruby classes so that the class is read in the same order as the operations will be performed leads to more legible code. Adding result objects enhances those same goals, and makes error conditions a first-class concern.