Derive local buckets from one scan

Four of a period's figures — the hour histogram, the busiest day, the per-day counts and the total — are all questions about the same rows of an append-only event table (local-listening-analytics), and all four are questions about local time. Written as SQL aggregates that is four scans, each needing a local-time conversion inside the engine. Written as one query returning the raw pairs, it is one scan and no conversion at all:

// adapted — the only query in this family that returns rows rather than a number
@Query("SELECT timestamp, listenedSecond FROM activity_event" +
       " WHERE timestamp BETWEEN :startTimestamp AND :endTimestamp")
suspend fun getSamplesInRange(startTimestamp: LocalDateTime, endTimestamp: LocalDateTime): List<Sample>

// adapted — one pass fills all four
val hours = IntArray(24)
val perDay = mutableMapOf<LocalDate, Int>()
var listened = 0L
samples.forEach { s ->
    hours[s.timestamp.hour]++
    perDay[s.timestamp.date] = (perDay[s.timestamp.date] ?: 0) + 1
    listened += s.listenedSecond
}
val busiest = perDay.maxByOrNull { it.value }

Traps

The engine's local-time modifier reads the process time zone. SQLite's 'localtime', and the equivalents elsewhere, resolve against whatever the OS hands the process — the TZ environment variable, a service account's zone, a container's default of UTC. That is not reliably the user's zone, it is not the same on a desktop build as on a phone, and it changes under you when the deployment does. The same query then returns a different histogram on two machines looking at identical data, with nothing to blame.

Four aggregates are four scans that can disagree. Each runs at its own moment; a write landing between the second and the third puts the total and the busiest day in different worlds. One scan is one consistent view of the period as well as being cheaper.

Only the bucketing moves — most aggregates still belong in SQL. The line is whether the grouping key is a local field. Distinct counts, top-N lists, first-ever minimums, joins and share-of-whole counts have no local component: they group by an id, or compare timestamps to each other on whatever scale both sides share. Those stay in the engine, where they cost one row of result instead of every row. Here, one range query of twelve returns raw rows; the other eleven are aggregates and are right to be.

The scan returns rows, so it grows with the window. A histogram query returns 24 numbers whatever the period; this returns one row per event. Project it down to the columns the derivation actually needs — two here — and know what the widest supported window costs before shipping a wider one. Where a period can be unbounded, this is the wrong shape.

The raw scan is also the emptiness check. Reading it first means an empty period can return early, before the seven aggregates this snapshot issues are sent at all — one round trip instead of eight, on exactly the periods a new user is looking at. Count what the snapshot issues, not what the family holds: the rest of it is issued by other loads on the same screen, which the return misses.

The samples must be typed so the ORM's converter still runs. Declaring the timestamp field as a raw number to "keep the projection cheap" hands back an encoding the derivation then has to interpret by hand, and the local hour is precisely the figure that goes wrong when it interprets it wrongly — see stored-timestamp-is-a-local-wall-clock. Ask for the date-time type; the hour and the date then read straight off it with no zone arithmetic anywhere in the path.

Do not push the local question into an index either. An index over a computed local field bakes one zone into stored bytes, and is silently wrong for every user in another one — and stale for the original user after a rule change. The zone belongs at the edge, applied once, in code.

Every new local bucket is a line in the existing loop, never a new query. This is where the pattern erodes: someone adds "busiest weekday" or "weekend versus weekday" and writes a fifth aggregate, because one more query looks smaller than touching working code. It is not — it is a fifth scan, a fifth chance to pick a zone, and the first of the answers that can disagree with the other four. The rule that keeps this stable is mechanical: if the grouping key comes off a local date-time, it is a line in the forEach.

Say so at the declaration. A query returning raw rows next to eleven returning numbers reads like someone who had not learned to aggregate. One comment naming the reason — local grouping, one scan, zone-independent — is what stops it being "optimised" into four GROUP BYs later.

Verifying it

1. Every local-time function in every query string. The correct outcome for this family is that the only hit is a comment saying why there is none:

   grep -rn --include='*.kt' -E "localtime|strftime\(" . | grep -v '/build/'
   

   One hit here, at the query that deliberately does not use it. Any hit inside an actual query string is a figure whose answer depends on where the process runs.

2. The shape of the range family — how many return rows, how many return numbers:

   grep -rn --include='*Dao*.kt' -E "suspend fun (get|query)[A-Za-z]*InRange" . | grep -v '/build/'
   

   Twelve here (32 unscoped, the rest being the datasource, repository and interface forwarding the same twelve). Read the return type of each: exactly one returns a List of raw samples, and it should be the one the local buckets are derived from. A second row-returning query in the same family usually means the same scan is being paid for twice.

3. Run the derivation against the engine, on the same data. Compute the hour histogram in code, then compute it in SQL with the engine's local-time modifier, and compare. They agree only when the process zone happens to equal the user's — so run it once with TZ=UTC and once with TZ=Asia/Ho_Chi_Minh in the environment. The code answer must not move; the SQL answer will.