by Ciro Santilli (@cirosantilli, 32)
Each transaction isolation level specifies what can or cannot happen when two queries are being run in parallel, i.e.: the memory semantics of the system.
Remember that queries can affects thousands of rows, and database systems like PostgreSQL can run multiple such queries at the same time.
Implementation specifics:
Example where this level is sufficient: nodejs/sequelize/raw/parallel_update_async.js.
nodejs/sequelize/raw/parallel_create_delete_empty_tag.js is an example which experimentally seems to be solved by REAPEATABLE READ, although we are not sure that this is truly the case and why. What is clear is that that example is not solved by the SQL READ COMMITTED isolation level.
In PostgreSQL, this is the first isolation level which can lead to postgreSQL serialization failures, this does not happen to SQL READ COMMITTED isolation level in that DBMS. You then have to retry the transaction.
nodejs/sequelize/raw/parallel_update_worker_threads.js contains a base example that can be used to test what can happen when queries are being run in parallel. But it is broken due to a sqlite3 Node.js package bug: github.com/mapbox/node-sqlite3/issues/1381...
nodejs/sequelize/raw/parallel_update_async.js is an async version of it. It should be just parallel enough to allow observing the same effects.
This is an example of a transaction where the SQL READ COMMITTED isolation level if sufficient.
These examples run queries of type: 
UPDATE "MyInt" SET i = i + 1
Sample execution: 
node --unhandled-rejections=strict ./parallel_update_async.js p 10 100
which does:
The fear then is that of a classic read-modify-write failure.
But as www.postgresql.org/docs/14/transaction-iso.html page makes very clear, including with an explicit example of type UPDATE accounts SET balance = balance + 100.00 WHERE acctnum = 12345;, that the default isolation level, SQL READ COMMITTED isolation level, already prevents any problems with this, as the update always re-reads selected rows in case they were previously modified.
If the first updater commits, the second updater will ignore the row if the first updater deleted it, otherwise it will attempt to apply its operation to the updated version of the row
Since in PostgreSQL "Read uncommitted" appears to be effectively the same as "Read committed", we won't be able to observe any failures on that database system for this example.
nodejs/sequelize/raw/parallel_create_delete_empty_tag.js contains an example where things can actually blow up in read committed.
This example is similar to nodejs/sequelize/raw/parallel_update_async.js, but now we are doing a separate SELECT, later followed by an update:
  • SELECT FROM to get i
  • update on Js code newI = i + 1
  • UPDATE SET the newI
Although this specific example is useless in itself, as we could just use UPDATE "MyInt" SET i = i + 1 as in nodejs/sequelize/raw/parallel_update_async.js, which automatically solves any concurrency issue, this kind of code could be required for example if the update was a complex function not suitably implemented in SQL, or if the update depends on some external data source.
Sample execution: 
node --unhandled-rejections=strict ./parallel_select_and_update.js p 2 10 'READ COMMITTED'
which does:
Another one: 
node --unhandled-rejections=strict ./parallel_select_and_update.js p 2 10 'READ COMMITTED' 'FOR UPDATE'
this will run SELECT FOR UPDATE rather than just SELECT
Observed behaviour under different SQL transaction isolation levels:
  • READ COMMITTED: fails. Nothing in this case prevents:
    • thread 1: SELECT, obtains i = 0
    • thread 2: SELECT, obtains i = 0
    • thread 2: newI = 1
    • thread 2: UPDATE i = 1
    • thread 1: newI = 1
    • thread 1: UPDATE i = 1
  • REPEATABLE READ: works. the manual mentions that if multiple concurrent updates would happen, only the first commit succeeds, and the following ones fail and rollback and retry, therefore preventing the loss of an update.
  • READ COMMITTED + SELECT FOR UPDATE: works. And does not do rollbacks, which probably makes it faster. With p 10 100, REPEATABLE READ was about 4.2s and READ COMMITTED + SELECT FOR UPDATE 3.2s on Lenovo ThinkPad P51 (2017).
    SELECT FOR UPDATE should be enough as mentioned at: www.postgresql.org/docs/13/explicit-locking.html#LOCKING-ROWS
    FOR UPDATE causes the rows retrieved by the SELECT statement to be locked as though for update. This prevents them from being locked, modified or deleted by other transactions until the current transaction ends. That is, other transactions that attempt UPDATE, DELETE, SELECT FOR UPDATE, SELECT FOR NO KEY UPDATE, SELECT FOR SHARE or SELECT FOR KEY SHARE of these rows will be blocked until the current transaction ends; conversely, SELECT FOR UPDATE will wait for a concurrent transaction that has run any of those commands on the same row, and will then lock and return the updated row (or no row, if the row was deleted). Within a REPEATABLE READ or SERIALIZABLE transaction, however, an error will be thrown if a row to be locked has changed since the transaction started. For further discussion see Section 13.4.
A non-raw version of this example can be seen at: nodejs/sequelize/parallel_select_and_update.js.
This example contains a deterministic demo of when postgreSQL serialization failures may happen.
Tested on PostgreSQL 13.5.
In this example, posts have tags. When a post is deleted, we check to see if there are now any empty tags, and now we want to delete any empty tags that the post deletion may have created.
If we are creating and deleting posts concurrently, a naive implementation might wrongly delete the tags of a newly created post.
This could be due to a concurrency issue of the following types.
Failure case 1:
  • thread 2: delete old post
  • thread 2: find all tags with 0 posts. Finds tag0 from the deleted old post which is now empty.
  • thread 1: create new post, which we want to have tag tag0
  • thread 1: try to create a new tag tag0, but don't because it already exists, this is done using SQLite's INSERT OR IGNORE INTO or PostgreSQL's INSERT ... ON CONFLICT DO NOTHING
  • thread 1: assign tag0 to the new post by adding an entry to the join table
  • thread 2: delete all tags with 0 posts. It still sees from its previous search that tag0 is empty, and deletes it, which then cascades into the join table
which would result in the new post incorrectly not having the tag0.
Failure case 2:
  • thread 2: delete old post
  • thread 2: find all tags with 0 posts
  • thread 1: create new post
  • thread 1: try to create a new tag tag0, but don't because it already exists
  • thread 2: delete all tags with 0 posts. It still sees from its previous search that tag0 is empty, and deletes it
  • thread 1: assign tag0 to the new post
which leads to a foreign key failure, because the tag does not exist anymore when the assignment happens.
Failure case 3:
  • thread 2: delete old post
  • thread 1: create new post, which we want to have tag tag0
  • thread 1: try to create a new tag tag0, and succeed because it wasn't present
  • thread 2: find all tags with 0 posts, finds the tag that was just created
  • thread 2: delete all tags with 0 posts, deleting the new tag
  • thread 1: assign tag0 to the new post
which leads to a foreign key failure, because the tag does not exist anymore when the assignment happens.
Sample executions:
  • node --unhandled-rejections=strict ./parallel_create_delete_empty_tag.js p 9 1000 'READ COMMITTED': PostgreSQL, 9 tags, DELETE/CREATE the tag0 test tag 1000 times, use READ COMMITTED
    Execution often fails, although not always. The failure is always: 
    error: insert or update on table "PostTag" violates foreign key constraint "PostTag_tagId_fkey"
    because the: 
    INSERT INTO "PostTag"
    tries to insert a tag that was deleted in the other thread, as it didn't have any corresponding posts, so this is the foreign key failure.
    TODO: we've never managed to observe the failure case in which tag0 is deleted. Is it truly possible? And if not, by which guarantee?
  • node --unhandled-rejections=strict ./parallel_create_delete_empty_tag.js p 9 1000 'READ COMMITTED' 'FOR UPDATE': do a SELECT ... FOR UPDATE before trying to INSERT.
    This is likely correct and the fastest correct method according to our quick benchmarking, about 20% faster than REPEATABLE READ.
    We are just now 100% sure it is corret becase we can't find out if the SELECT in the DELETE subquery could first select some rows, which are then locked by the tag creator, and only then locked by DELETE after selection. Or does it re-evaludate the SELECT even though it is in a subquery?
  • node --unhandled-rejections=strict ./parallel_create_delete_empty_tag.js p 9 1000 'REPEATABLE READ': repeatable read
    We've never observed any failures with this level. This should likely fix the foreign key issue according to the PostgreSQL docs, since:
    • the DELETE "Post" commit cannot start to be seen only in the middle of the thread 1 transaction
    • and then if DELETE happened, the thread 1 transaction will detect it, ROLLBACK, and re-run. TODO how does it detect the need rollback? Is it because of the foreign key? It is very hard to be sure about this kind of thing, just can't find the information. Related: postgreSQL serialization failure.
  • node --unhandled-rejections=strict ./parallel_create_delete_empty_tag.js p 9 1000 'SERIALIZABLE': serializable
  • node --unhandled-rejections=strict ./parallel_create_delete_empty_tag.js p 9 1000 'NONE': magic value, don't use any transaction. Can blow up of course, since even less restrictions than READ COMMITTED
All executions use 2 threads.
Some theoretical notes:
  • Failure case 3 is averted by a READ COMMITTED transaction, because thread 2 won't see the uncommitted tag that thread 1 created, and therefore won't be able to delete it
stackoverflow.com/questions/10935850/when-to-use-select-for-update from SELECT FOR UPDATE also talks about a similar example, and has relevant answers.

Tagged

Ancestors

  1. SQL transaction
  2. SQL feature
  3. SQL
  4. Relational database management system
  5. Relational database
  6. Database
  7. Software
  8. Computer
  9. Information technology
  10. Area of technology
  11. Technology
  12. Index

Incoming links

Synonyms

Discussion (0)

Sign up or sign in create discussions.

There are no discussions about this article yet.

View article source