This has been a hair-pulling rabbit hole of an issue. #1931 and others.
When the `next-campaign-subscribers` query that fetches $n subscribers
per batch for a campaign returns no results, the manager assumes
that the campaign is done and marks as finished.
Marathon debugging revealed fundamental flaws in qyery's logic that
would incorrectly return 0 rows under certain conditions.
- Based on the "layout" of subscribers for eg: a series of blocklisted
subscribers between confirmed subscribers.
A series of unconfirmed subscribers in a batch belonging to a double
opt-in list.
- Bulk import blocklisting users, but not marking their subscriptions
as 'unsubscribed'.
- Conditions spread across multiple CTEs resulted in returning an
arbitrary number of rows and $N per batch as the selected $N rows
would get filtered out elsewhere, possibly even becoming 0.
After fixing this and testing it on our prod instance that has
15 million subscribers and ~70 million subscriptions in the
`subscriber_lists` table, ended up discovered significant inefficiences
in Postgres query planning. When `subscriber_lists` and campaign list IDs
are joined dynamically (CTE or ANY() or any kind of JOIN that involves)
a query, the Postgres query planner is unable to use the right indexes.
After testing dozens of approaches, discovered that statically passing
the values to join on (hardcoding or passing via parametrized $1 vars),
the query uses the right indexes. The difference is staggering.
For the particular scenario on our large prod DB to pull a batch,
~15 seconds vs. ~50ms, a whopping 300x improvement!
This patch splits `next-campaign-subscribers` into two separate queries,
one which fetches campaign metadata and list_ids, whose values are then
passed statically to the next query to fetch subscribers by batch.
In addition, it fixes and refactors broken filtering and counting logic
in `create-campaign` and `next-campaign` queries.
Closes#1931, #1993, #1986.
This commit splits roles into two, user roles and list roles, both of which
are attached separately to a user.
List roles are collection of lists each with read|write permissions, while
user roles now have all permissions except for per-list ones.
This allows for easier management of roles, eliminating the need to clone and
create new roles just to adjust specific list permissions.
- Filter lists by permitted list IDs in DB get calls.
- Split getLists() handlers into two (one, all) for clarity.
- Introduce new `subscribers:get_by_list` permission.
- Tweak UI rendering to work with new per-list permssions.
- Add materialized views for list -> subscriber counts, dashboard chart,
and dashboard aggregate stats that slow down significantly on large
databases (with millions or tens of millions of subscribers). These
slow queries involve full table scan COUNTS().
- Add a toggle to enable caching slow results in Settings -> Performance.
- Add support for setting a cron string that crons and periodically
refreshes aggregated stats in materialized views.
Closes#1019.
- Sent count is no longer the batch size fetched from the DB but is
the actual count of messages sent.
- Pausing and resuming now accurately tracks the last subscriber that
was processed and resumes from there.
- Fix multiple concurrent campaigns blocking.
Closes#1616. Closes#905. Closes#1496. Closes#1250. Closes#1010.
- Add new 'Subscriptions' table on the subscriber list form that shows subs,
IP, and other data.
- Add new `meta` JSONB field to `subscriber_lsts` table.
Closes#1329.
