From 94feb2ba20994171ea40e17c31c262c90379251e Mon Sep 17 00:00:00 2001
From: mindsers <nathanael@cherrier.dev>
Date: Sun, 23 Aug 2026 22:27:33 +0200
Subject: [PATCH] feat(territories): split the legacy campaign backfill into
 per-period campaigns
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

The campaigns migration collapsed all legacy type='campaign' attributions
into one synthetic ended 'Campagne' per congregation. This follow-up data
migration splits each synthetic into one campaign per date cluster (a gap of
more than 30 days between an attribution's start and the latest end seen so
far opens a new cluster), named by its period ('Campagne 03/2024'),
repoints the attributions and drops the emptied synthetic. Synthetics are
recognized by the backfill's exact stamps (name + activatedAt = startDate +
endedAt = endDate), so campaigns that went through the real lifecycle —
even ones named 'Campagne' — and single-cluster synthetics are untouched.
Single data-modifying CTE chain (temp tables don't survive Prisma's
statement execution); verified against seeded multi-cluster, decoy and
single-cluster states on the dev database.
---
 .../migration.sql                             | 95 +++++++++++++++++++
 1 file changed, 95 insertions(+)
 create mode 100644 app/database/migrations/20260823200000_split_legacy_campaign_backfill/migration.sql

diff --git a/app/database/migrations/20260823200000_split_legacy_campaign_backfill/migration.sql b/app/database/migrations/20260823200000_split_legacy_campaign_backfill/migration.sql
new file mode 100644
index 00000000..d8fc10a7
--- /dev/null
+++ b/app/database/migrations/20260823200000_split_legacy_campaign_backfill/migration.sql
@@ -0,0 +1,95 @@
+-- Follow-up to 20260823000000_add_publishing_campaigns: that backfill collapsed
+-- every congregation's legacy `type = 'campaign'` attributions into ONE synthetic
+-- ended campaign named 'Campagne'. Real history usually spans several distinct
+-- drives, so split each synthetic campaign into one campaign per date cluster:
+-- a gap of more than 30 days between an attribution's start and the latest end
+-- seen so far opens a new cluster. Synthetic campaigns are recognized by the
+-- exact stamps the backfill wrote (name 'Campagne', activatedAt = startDate,
+-- endedAt = endDate) — a campaign that went through the real lifecycle never
+-- has activation/end timestamps exactly equal to its day-granular bounds.
+-- Single-cluster synthetics are left untouched (splitting would be a rename).
+-- One data-modifying CTE chain: temp tables don't survive Prisma's statement
+-- execution.
+
+WITH synth AS (
+    SELECT id, "congregationId"
+    FROM "Campaign"
+    WHERE name = 'Campagne'
+      AND "activatedAt" = "startDate"
+      AND "endedAt" = "endDate"
+),
+attrs AS (
+    SELECT a.id,
+           a."congregationId",
+           a."campaignId",
+           a."startDate",
+           COALESCE(a."endDate", a."startDate") AS eff_end
+    FROM "Attribution" a
+    JOIN synth s ON s.id = a."campaignId"
+),
+marked AS (
+    SELECT attrs.*,
+           CASE
+               WHEN "startDate" > COALESCE(
+                   MAX(eff_end) OVER (
+                       PARTITION BY "campaignId"
+                       ORDER BY "startDate", id
+                       ROWS BETWEEN UNBOUNDED PRECEDING AND 1 PRECEDING
+                   ),
+                   "startDate"
+               ) + INTERVAL '30 days'
+               THEN 1 ELSE 0
+           END AS opens_cluster
+    FROM attrs
+),
+clustered AS (
+    SELECT marked.*,
+           SUM(opens_cluster) OVER (PARTITION BY "campaignId" ORDER BY "startDate", id) AS cluster
+    FROM marked
+),
+-- Only campaigns that actually split into several clusters are rewritten.
+bounds AS (
+    SELECT "campaignId",
+           "congregationId",
+           cluster,
+           MIN("startDate") AS cluster_start,
+           MAX(eff_end)     AS cluster_end
+    FROM clustered
+    WHERE "campaignId" IN (
+        SELECT "campaignId" FROM clustered GROUP BY "campaignId" HAVING COUNT(DISTINCT cluster) > 1
+    )
+    GROUP BY "campaignId", "congregationId", cluster
+),
+-- One already-ended campaign per cluster, named by its period. Cluster bounds
+-- within one congregation are disjoint (30-day gaps) and there is only one
+-- synthetic campaign per congregation, so (congregationId, startDate, endDate)
+-- uniquely identifies the created row for the repointing below.
+created AS (
+    INSERT INTO "Campaign" ("name", "startDate", "endDate", "activatedAt", "endedAt", "congregationId", "updatedAt")
+    SELECT 'Campagne ' || to_char(b.cluster_start, 'MM/YYYY'),
+           b.cluster_start,
+           b.cluster_end,
+           b.cluster_start,
+           b.cluster_end,
+           b."congregationId",
+           CURRENT_TIMESTAMP
+    FROM bounds b
+    RETURNING id, "congregationId", "startDate", "endDate"
+),
+repointed AS (
+    UPDATE "Attribution" a
+    SET "campaignId" = c.id
+    FROM clustered lc
+    JOIN bounds b
+      ON b."campaignId" = lc."campaignId" AND b.cluster = lc.cluster
+    JOIN created c
+      ON c."congregationId" = b."congregationId"
+     AND c."startDate" = b.cluster_start
+     AND c."endDate" = b.cluster_end
+    WHERE a.id = lc.id
+    RETURNING lc."campaignId" AS old_campaign_id
+)
+-- The split synthetics are now empty (every attribution was repointed; the
+-- backfill never paused anything) — drop them.
+DELETE FROM "Campaign"
+WHERE id IN (SELECT DISTINCT old_campaign_id FROM repointed);