Let’s continue to dive in PostgreSQL Concurrency. Last week’s article PostgreSQL Concurrency: Isolation and Locking was a primer on PostgreSQL isolation and locking properties and behaviors.

Today’s article takes us a step further and builds on what we did last week, in particular the database modeling for a tweet like application. After having had all the characters from Shakespeare’s A Midsummer Night’s Dream tweet their own lines in our database in PostgreSQL Concurrency: Data Modification Language, it’s time for them to do some actions on the tweets: likes and retweet.

Of course, we’re going to put concurrency to the test, so we’re going to have to handle very very popular tweets from the play!

Modeling for Concurrency

We should have another modeling pass on the tweet.message table now. With what we learned about concurrency in PostgreSQL, it’s easy to see that we won’t get anywhere with the current model. Remember when Donald Knuth said

We should forget about small efficiencies, say about 97% of the time: premature optimization is the root of all evil. Yet we should not pass up our opportunities in that critical 3%.

Database systems have been designed to handle concurrency so that your application’s code doesn’t have to. One part for the critical 3% is then related to concurrent operations, and the one that is impossible to implement in a both fast and correct way is a concurrent update on the same target row.

In our model here, given how the application works, we know that messages will get concurrent update activity for the favs and rts counters. So while the previous model looks correct with respect to normal forms — the counters are dependent on the message’s key — we know that concurrent activity is going to be hard to handle in production.

So here’s a smarter version of the activity parts of the database model:

begin;

create type tweet.action_t
    as enum('rt', 'fav', 'de-rt', 'de-fav');

create table tweet.activity
 (
  id          bigserial primary key,
  messageid   bigint not null references tweet.message(messageid),
  datetime    timestamptz not null default now(),
  action      tweet.action_t not null,

  unique(messageid, datetime, action)
 );

commit;

In this version, the counters have disappeared, replaced by a full record of the base information needed to compute them. We now have an activity list with a denormalized ENUM for possible actions.

To get the rts and favs counters back from this schema, we count lines in the activity records associated with a given messageid:

  select   count(*) filter(where action = 'rt')
         - count(*) filter(where action = 'de-rt')
         as rts,
           count(*) filter(where action = 'fav')
         - count(*) filter(where action = 'de-fav')
         as favs
    from tweet.activity
         join tweet.message using(messageid)
   where messageid = :id;

Reading the current counter value has become quite complex when compared to just adding a column to your query output list. On the other hand, when adding a rt or a fav action to a message, we transform the SQL:

update tweet.message set rts = rts +1 where messageid = :id;

This is what we use instead:

insert into tweet.activity(messageid, action) values(:id, 'rt');

The reason why replacing an update with an insert is interesting is concurrency behavior and locking. In the first version, retweeting has to wait until all concurrent retweets are done, and the business model wants to sustain as many concurrent activities on the same small set of messages as possible (read about influencer accounts).

The insert has no concurrency because it targets a row that doesn’t exist yet. We register each action into its own tuple and require no locking to do that, allowing our production setup of PostgreSQL to sustain a much larger load.

Now, computing the counters each time we want to display them is costly. And the counters are displayed on every tweet message. We need a way to cache that information, and we’ll see about that in a follow-up article about Computing and Caching in SQL.

Putting Concurrency to the Test

When we benchmark the concurrency properties of the two statements above, we quickly realize that the activity table is badly designed. The unique constraint includes a timestamptz field, which in PostgreSQL is only precise down to the microsecond.

This kind of made-up unique constraint means we now have these errors to deal with:

Error: Database error 23505: duplicate key value violates unique     ⏎
  constraint "activity_messageid_datetime_action_key"
DETAIL: Key (messageid, datetime, action)                            ⏎
           =(2, 2017-09-19 18:00:03.831818+02, rt) already exists.

The best course of action here is to do this:

    alter table tweet.activity
drop constraint activity_messageid_datetime_action_key;

Now we can properly compare the concurrency scaling of the insert and the update based version. In case you might be curious about it, here’s the testing code that’s been used:

(defpackage #:concurrency
  (:use #:cl #:appdev)
  (:import-from #:lparallel
                #:*kernel*
                #:make-kernel #:make-channel
                #:submit-task #:receive-result
                #:kernel-worker-index)
  (:import-from #:cl-postgres-error
                #:database-error)
  (:export      #:*connspec*
                #:concurrency-test))

(in-package #:concurrency)

(defparameter *connspec* '("appdev" "dim" nil "localhost"))

(defparameter *insert-rt*
  "insert into tweet.activity(messageid, action) values($1, 'rt')")

(defparameter *update-rt*
  "update tweet.message set rts = coalesce(rts, 0) + 1 where messageid = $1")

(defun concurrency-test (workers retweets messageid
                         &optional (connspec *connspec*))
  (format t "Starting benchmark for updates~%")
  (with-timing (rts seconds)
      (run-workers workers retweets messageid *update-rt* connspec)
    (format t "Updating took ~f seconds, did ~d rts~%" seconds rts))

  (format t "~%")

  (format t "Starting benchmark for inserts~%")
  (with-timing (rts seconds)
      (run-workers workers retweets messageid *insert-rt* connspec)
    (format t "Inserting took ~f seconds, did ~d rts~%" seconds rts)))

(defun run-workers (workers retweets messageid sql
                    &optional (connspec *connspec*))
  (let* ((*kernel* (lparallel:make-kernel workers))
         (channel  (lparallel:make-channel)))
    (loop repeat workers
       do (lparallel:submit-task channel #'retweet-many-times
                                 retweets messageid sql connspec))

    (loop repeat workers sum (lparallel:receive-result channel))))

(defun retweet-many-times (times messageid sql
                           &optional (connspec *connspec*))
  (pomo:with-connection connspec
    (pomo:query
     (format nil "set application_name to 'worker ~a'"
             (lparallel:kernel-worker-index)))
    (loop repeat times sum (retweet messageid sql))))

(defun retweet (messageid sql)
  (handler-case
      (progn
        (pomo:query sql messageid)
        1)
    (database-error (c)
      (format t "Error: ~a~%" c)
      0)))

Here’s a typical result with a concurrency of 100 workers all wanting to do 10 retweet in a loop using a messageid, here message 3. While it’s not representative to have them loop 10 times to retweet the same message, it should help create the concurrency effect we want to produce, which is having several concurrent transactions waiting in turn in order to have a lock access to the same row.

The theory says that those concurrent users will have to wait in line, and thus spend time waiting for a lock on the PostgreSQL server. We should see that in the timing reports as a time difference:

CL-USER> (concurrency::concurrency-test 100 10 3)
Starting benchmark for updates
Updating took 3.099873 seconds, did 1000 rts

Starting benchmark for inserts
Inserting took 2.132164 seconds, did 1000 rts

The update variant of the test took almost 50% as much time to complete than the insert variant, with this level of concurrency. Given that we have really simple SQL statements, we can attribute the timing difference to having had to wait in line. Basically, the update version spent almost 1 second out of 3 seconds waiting for a free slot.

In another test with even more concurrency pressure at 50 retweets per worker, we can show that the results are repeatable:

CL-USER> (concurrency::concurrency-test 100 50 6)
Starting benchmark for updates
Updating took 5.070135 seconds, did 5000 rts

Starting benchmark for inserts
Inserting took 3.739505 seconds, did 5000 rts

This article is extracted from my book The Art of PostgreSQL, which teaches SQL to developers so that they may replace thousands of lines of code with very simple queries. The book has a full chapter about Data Manipulation and Concurrency Control in PostgreSQL, including caching with materialized views, check it out!


Conclusion

If you know that your application has to scale, think about how to avoid concurrent activity that competes against a single shared resource. Here, this shared resource is the rts field of the tweet.message row that you target, and the concurrency behavior is going to be fine if the retweet activity is well distributed. As soon as many users want to retweet the same message, then the update solution has a non-trivial scalability impact.

Now, we’re going to implement the tweet.activity based model. In this model, the number of retweets needs to be computed each time we display it, and it’s part of the visible data.

Also, in the general case, it’s impossible for our users to know for sure how many retweets have been made so that we can implement a cache with eventual consistency properties in the next article of our series about concurrency in PostgreSQL.