Stats: Active accounts by payment level

[pauamma]

One of the stats we need to provide is active accounts by account level (free, paid, premium, seed). Currently, the active accounts come from a per-cluster table with the user id as primary key and the time of last activity, in which users are added/updated whenever they do something that makes them active.

DW::StatData::ActiveAccounts handle collecting those stats, basically running something like the following request to retrieve the number of accounts active in the last 30 days:

SELECT COUNT(*) FROM clustertrack2 WHERE timeactive > UNIX_TIMESTAMP()-30*86400

That table doesn't include the account level, because (so far) it doesn't need to. This means that to break down active accounts by account level, I'd need to retrieve the complete list of active accounts and check the account level for each, which is much more costly than the current SQL.

To avoid that, the obvious solution is to add the account level to that table. However, this could cause discrepancies if the account level changes after the latest activity. For instance, if a paid account expires after the last activity, it would still be counted as "active paid" in the breakdown of active accounts, but free in the accounts by level stats, possibly leading to more accounts counted as "active paid" than as "paid". The only way to fix that is to have the account level transitions (paid account expiring, payment received, etc...) "correct" the table, but that gets rather ugly as the number of stats that needs a breakdown by account level goes up (and may not be possible for some stats at all, depending on how they're collected). Also, it can lead to biased data. For instance, if a free account gets paid for the first time after the last activity but before the stats are collected, it will be counted as paid even though all its activity was as a free account.

Adding the account level to the primary key (and thus, keeping separate activity records for each user id) isn't feasible, because that field would be null for existing activity records, which makes it unsuitable for a primary key. In addition, that wouldn't really help, because we'd still need to choose between activity records for the same account, lest we count it twice (or more). If having (according to stats) more active paid accounts than paid accounts (or than active accounts) makes stats look unreliable, think how much worse it would be to have more active paid accounts than total accounts. (Can you hear the "DW is inflating its stats" cries from here? :-) )

Anyway: the best compromise (IMO) seems to capture the account level in the activity table at the time that table is updated, and to leave it unchanged afterward. Opinions? Questions? Blindingly obvious solutions that I missed?

Comments

[pauamma]

See http://hg.dwscoalition.org/dw-free/file/83ed7fefd9d5/bin/upgrading/update-db-general.pl. Also, grep -r the source tree for the name of the table should point you to the module that wraps that table, if applicable. That module, in turn, may provide further insight on how the table is (meant to be) used.

Re your general idea: that may work, if the performance isn't intolerably low. (Need to look at that in more depth, unless mark or one of the ops people already knows.) Never mind. That won't work, because clustertrack2 is clustered and the account level history table (IIRC) isn't, so we can't join them (different database servers).

[shirokage]

Shouldn't we use the active_user table instead of the clustertrack2?  After all, clustertrack2 tracks where the users are active (ie. which server cluster they're on), but we don't need to know where they are, do we?  We only need to know how many active paid users there are at the time when the stats are taken.  Assuming that the time attribute in the active_user table (see below) is the time the user was last active, we can join it with the dw_paidstatus table (assuming this is the account history table, because this is the only one I saw that kinda fit the description).

The active_user table:
CREATE TABLE active_user (
    userid INT UNSIGNED NOT NULL,
    type CHAR(1) NOT NULL,
    time INT UNSIGNED NOT NULL,
    KEY (userid),
    KEY (time)
)

And the dw_paidstatus table:
CREATE TABLE dw_paidstatus (
    userid int unsigned NOT NULL,
    typeid smallint unsigned NOT NULL,
    expiretime int unsigned,
    permanent tinyint unsigned NOT NULL,
    lastemail int unsigned,
    PRIMARY KEY (userid),
    INDEX (expiretime)
)

So if the assumption about the time attribute in the active_user table applies, then this:
SELECT  COUNT(*)
FROM    active_user au JOIN dw_paidstatus dwps ON au.memberID = dwps.memberID
WHERE   dwps.typeid = (whatever the paid IDs are, several of them can be allowed using OR statements) AND
        au.time > dwps.beginTime AND
        au.time < dwps.expireTime AND
        dwps.beginTime < getDate() AND
        dwps.expireTime > getDate();
should work if we add a column to contain when that account level period (beginTime) started to the dw_paidstatus table, and that the records were kept by adding a new row for every account level change.  But then, we'll need to change the primary key as well, because one user may have several account level changes, thus several entries, and the primary key needs to be unique.  (This method will generate more entries than the existing method, but the new model will allow us to track and analyze trends like when people are buying paid accounts and etc.)

This might work?  Maybe?  I hope I'm not doing something completely stupid...

[pauamma]

Hmm. IIRC my reasoning then, I wanted to get rid eventually of one of those tables eventually. But I don't remember now why I chose to keep clustertrack2.

[shirokage]

Maybe to keep track of server cluster usage and load balancing?

[pauamma]

That's a good reason by itself. :-) But thinking back of it, ISTR I meant to get rid of bin/maint/stats*.pl eventually and to move all stats to the DW::Stats framework.