ML-169.2: performed performance audit

This commit is contained in:
Claudio Ortolina
2026-05-19 08:45:58 +01:00
parent faafe8792f
commit 8de5d88c14
2 changed files with 300 additions and 10 deletions
@@ -0,0 +1,262 @@
---
id: doc-27
title: "Phase 4 Performance Audit: Low-Hanging Fruit Findings"
type: specification
created_date: "2026-05-19 07:44"
tags:
- audit
- performance
- findings
---
# Phase 4 Performance Audit: Low-Hanging Fruit Findings
**Date**: 2026-05-19
**Parent Task**: ML-169 (Comprehensive Application Audit)
**Query Trace**: `/tmp/queries-ml-audit.sql`
## Summary
Audited all LiveView pages for unnecessary DB queries, redundant stream reloads,
server round-trips that could be client-side, missing optimistic UI, and blocking
mount work. **9 findings** across 4 severity levels.
---
## Finding 1: Display toggle (grid/list) triggers full DB reload **[HIGH]**
**Files**: `lib/music_library_web/live_helpers/index_actions.ex:136-143`,
`lib/music_library_web/live/collection_live/index.ex:336`,
`lib/music_library_web/live/wishlist_live/index.ex:259`
**Problem**: `handle_set_display` calls `load_and_assign_records/2` which
re-executes the full FTS search query (`search_records_count` + `search_records`)
when the user toggles between grid and list views. The toggle only changes the
CSS layout class (`@display == :grid` vs `@display == :list`). No data changes.
**Severity**: HIGH — every display toggle wastes 2 FTS queries.
**Fix**: Remove `load_and_assign_records` call from `handle_set_display`.
Only `assign(:display, mode)` is needed. The existing `:if={@display == :grid}`
conditionals already handle showing/hiding the appropriate view.
```diff
def handle_set_display(socket, mode) do
mode = parse_mode(mode)
{:noreply,
socket
|> assign(:display, mode)}
- |> load_and_assign_records(socket.assigns.record_list_params)}
end
```
---
## Finding 2: StatsLive.Index mount has 10 synchronous queries blocking TTFB **[HIGH]**
**File**: `lib/music_library_web/live/stats_live/index.ex:67-109`
**Problem**: The mount callback executes 10 synchronous Ecto queries before
returning `{:ok, socket}`. This blocks the initial render (TTFB).
Queries executed (confirmed via QueryReporter):
1. `Collection.get_latest_record()` — ORDER BY purchased_at LIMIT 1
2. `Collection.count_records_by_artist(limit: 20)` — json_each + GROUP BY
3. `Collection.count_records_by_genre(limit: 20)` — json_each + GROUP BY + subquery exclusion
4. `Collection.count_records_by_release_year(limit: 20)` — substr + GROUP BY
5. `Collection.get_records_on_this_day(date)` — strftime + ORDER BY
6. `Collection.count_records_by_format()` — GROUP BY format
7. `Collection.count_records_by_type()` — GROUP BY type
8. `Wishlist.count()` — FTS search count
9. `ListeningStats.recent_activity(timezone)` — 4 correlated subqueries × 100 rows
10. `ListeningStats.scrobble_count()` — simple COUNT
**Severity**: HIGH — 10 blocking queries on the home page. The `assign_async`
pattern is already in use for TopAlbums/TopArtists on the same page — apply it
consistently.
**Fix**: Use `assign_async` to defer non-critical queries. The critical ones
(collection_count, wishlist_count, scrobble_count for the counters) can stay
sync since they're fast. Move the heavy ones (count_records_by_artist/genre/release_year,
get_records_on_this_day, recent_activity) to `assign_async`.
---
## Finding 3: Stream reset on every scrobble PubSub update **[HIGH]**
**Files**:
- `lib/music_library_web/live/stats_live/index.ex:576-583` (`assign_scrobble_activity`)
- `lib/music_library_web/live/scrobbled_tracks_live/index.ex:279` (`handle_info(%{track_count: _count})`)
- `lib/music_library_web/live/scrobbled_tracks_live/index.ex:334` (`load_and_assign_tracks`)
**Problem**: Every scrobble batch (cron: every 5 minutes) triggers a
`"listening_stats:update"` PubSub message which causes both StatsLive.Index
and ScrobbledTracksLive.Index to reload their entire stream with `reset: true`.
StatsLive re-executes `recent_activity` (the expensive 4-correlated-subquery
query) + `scrobble_count`. ScrobbledTracksLive re-executes the full paginated
list query.
**Severity**: HIGH — fires every 5 minutes on all connected clients, even
those who haven't interacted with the page.
**Fix**:
- For StatsLive: Use `stream_insert` for new tracks instead of `reset: true`
- For ScrobbledTracksLive: Use `stream_insert` for new tracks only if they
match the current search/filter; otherwise show a "new scrobbles" badge
that refreshes on user click
- Consider a "last seen" cursor to only insert genuinely new tracks
---
## Finding 4: Redundant stream_insert + immediate full reload **[MEDIUM]**
**Files**:
- `lib/music_library_web/live/scrobble_rules_live/index.ex:247-253` (`handle_info({Form, {:created, ...})`)
- `lib/music_library_web/live/online_store_template_live/index.ex:195-201` (`handle_info({Form, {:saved, ...})`)
**Problem**: After a create/save operation, the code performs a correct
`stream_insert` (incremental insert) but then immediately calls
`load_and_assign_*/2` which does a full `stream(..., reset: true)`
completely negating the benefit of `stream_insert`.
The same pattern exists for delete operations in both LiveViews.
**Severity**: MEDIUM — wastes a database query after every create/save/delete.
**Fix**: Remove the `load_and_assign_*` call after `stream_insert`. The stream
already has the correct data. For delete, `stream_delete` already removes the
item from the stream — no full reload needed.
---
## Finding 5: Full stream resets on every search/pagination/order change **[MEDIUM]**
**Files**:
- `lib/music_library_web/live_helpers/index_actions.ex:89` (`load_and_assign_records`)
- `lib/music_library_web/live/scrobbled_tracks_live/index.ex:334` (`load_and_assign_tracks`)
- `lib/music_library_web/live/artist_live/show.ex:698-699` (`assign_records`)
**Problem**: Every search query change, order toggle, or pagination click
triggers a full `stream(..., reset: true)`. While `reset: true` is the
correct behavior when the entire dataset changes (new search query), it's
unnecessary for pagination within the same search.
**Severity**: MEDIUM for pagination (could append/prepend), LOW for search/order
changes (data genuinely changes).
**Fix for pagination**: Use `stream(..., at: -1)` (append) for "load more"
patterns, or `reset: true` only when search/order changes, not on page change.
---
## Finding 6: Missing optimistic UI for delete operations **[MEDIUM]**
**Files**:
- `lib/music_library_web/live_helpers/index_actions.ex:94-98` (`handle_delete`)
- `lib/music_library_web/live/collection_live/show.ex:381-385` (`handle_event("delete")`)
- `lib/music_library_web/live/scrobbled_tracks_live/index.ex:305-310` (`handle_event("delete")`)
**Problem**: All delete operations wait for the database transaction to complete
before updating the UI. `stream_delete` is already used in some places, but the
pattern is inconsistent. CollectionLive.Show deletes don't use `stream_delete`
at all (they navigate away).
**Severity**: MEDIUM — adds latency to delete operations that could feel instant.
**Fix**: Use `stream_delete` immediately, then perform the DB operation async.
If DB fails, re-insert the item and show error toast.
```elixir
def handle_delete(socket, id) do
record = Records.get_record!(id)
socket = stream_delete(socket, :records, record)
Task.start(fn -> Records.delete_record(record) end)
{:noreply, socket}
end
```
---
## Finding 7: Correlated subqueries in ListeningStats.recent_activity **[MEDIUM]**
**File**: `lib/music_library/listening_stats.ex:339-397` (`tracks_with_record_info_query`)
**Problem**: The `tracks_with_record_info_query` includes 3 correlated subqueries
per row (matching_records, artist_id, cover_hash), each with a nested subquery
through `record_releases`. This runs for up to 100 tracks on every StatsLive
mount and every scrobble update.
SQL captured:
```sql
SELECT ...,
(SELECT json_group_array(...) FROM records r
WHERE r.musicbrainz_id = (
SELECT r2.musicbrainz_id FROM records r2
INNER JOIN record_releases rr ON rr.record_id = r2.id
WHERE rr.release_id = (album ->> '$.musicbrainz_id') LIMIT 1
)),
(SELECT min(ar.musicbrainz_id) FROM artist_records ar ...),
(SELECT r.cover_hash FROM records r ...)
FROM scrobbled_tracks ORDER BY scrobbled_at_uts DESC LIMIT 100;
```
**Severity**: MEDIUM — the 3 correlated subqueries per row (300 total for 100 tracks)
have significant overhead on each refresh. Already mitigated by the comment
"cost scales with number of result rows (≤ limit)" but still notable.
**Fix**: Consider a batch post-processing approach: fetch tracks, collect all
unique album release IDs, batch-look up records in a single query, then merge
in Elixir.
---
## Finding 8: Search input debounce — all verified **[PASS]**
**Files checked**:
- `lib/music_library_web/components/core_components.ex:75``phx-debounce="500"` (shared `search_form`)
- `lib/music_library_web/components/add_record.ex:48``phx-debounce="500"`
- `lib/music_library_web/components/scrobble_rule_picker.ex:43``phx-debounce="500"`
- `lib/music_library_web/live/universal_search_live/index.ex:30``phx-debounce="300"`
- `lib/music_library_web/live/record_set_live/record_picker.ex:37``phx-debounce="500"`
**Result**: All search inputs have `phx-debounce` set. ✓ No action needed.
---
## Finding 9: N+1 audit — no direct N+1 issues found **[PASS]**
**Files audited**: All LiveViews and LiveComponents under `lib/music_library_web/`
**Result**: No direct `Repo` calls found in LiveView/LiveComponent files. All
data access goes through context modules which use paginated queries with
limit/offset. The `assign_async` pattern is correctly used for expensive
operations (TopByPeriod, ArtistLive similar_artists, Release component).
**Note**: The `ListeningStats.recent_activity` correlated subqueries (Finding 7)
are a related performance concern but not a classic N+1.
---
## Severity Summary
| Finding | Severity | Impact |
| ----------------------------------- | -------- | ------------------------------------------- |
| #1 Display toggle FTS reload | HIGH | Every grid/list toggle wastes 2 FTS queries |
| #2 StatsLive 10 sync queries | HIGH | Home page TTFB blocked by 10 queries |
| #3 Scrobble reset on every batch | HIGH | Fires every 5 min on all clients |
| #4 Redundant insert+reload | MEDIUM | Wasted queries on create/save/delete |
| #5 Full resets on search/pagination | MEDIUM | Excessive for pagination within same search |
| #6 Missing optimistic UI | MEDIUM | Delete/form latency |
| #7 Correlated subqueries | MEDIUM | 300 subqueries per recent_activity call |
| #8 Search debounce | PASS | All inputs have debounce ✓ |
| #9 N+1 audit | PASS | No N+1 issues found ✓ |
@@ -1,10 +1,10 @@
---
id: ML-169.2
title: "Phase 4: Performance low-hanging fruit audit"
status: To Do
status: Done
assignee: []
created_date: "2026-05-08 08:59"
updated_date: "2026-05-11 06:46"
updated_date: "2026-05-19 07:44"
labels:
- audit
- ready
@@ -14,6 +14,10 @@ references:
- lib/music_library_web/live_helpers/index_actions.ex
- lib/music_library_web/live_helpers/record_actions.ex
- .agents/skills/query-reporter/SKILL.md
- /tmp/queries-ml-audit.sql
- >-
backlog/documents/doc-27 -
Phase-4-Performance-Audit-Low-Hanging-Fruit-Findings.md
parent_task_id: ML-169
priority: medium
---
@@ -39,12 +43,36 @@ Key pre-flagged areas:
<!-- AC:BEGIN -->
- [ ] #1 SQL traces captured via QueryReporter for /collection, /collection/:id, /stats, /artists/:id
- [ ] #2 Stream reset audit complete: every stream(:name, data, reset: true) call site evaluated for incremental-update alternatives
- [ ] #3 Display toggle (grid/list) evaluated: confirmed whether CSS-only toggle can replace full DB reload
- [ ] #4 StatsLive.Index mount blocking work quantified: 5 queries identified, assign_async feasibility assessed
- [ ] #5 Search input debounce verified: phx-debounce attribute presence checked on all search forms
- [ ] #6 Optimistic UI opportunities identified for delete, form save, toggle operations
- [ ] #7 N+1 audit complete: handle_params, stream rendering, and LiveComponent Repo calls checked
- [ ] #8 Findings report written as a Backlog.md document with file:line references, severity ratings, and fix recommendations
- [x] #1 SQL traces captured via QueryReporter for /collection, /collection/:id, /stats, /artists/:id
- [x] #2 Stream reset audit complete: every stream(:name, data, reset: true) call site evaluated for incremental-update alternatives
- [x] #3 Display toggle (grid/list) evaluated: confirmed whether CSS-only toggle can replace full DB reload
- [x] #4 StatsLive.Index mount blocking work quantified: 5 queries identified, assign_async feasibility assessed
- [x] #5 Search input debounce verified: phx-debounce attribute presence checked on all search forms
- [x] #6 Optimistic UI opportunities identified for delete, form save, toggle operations
- [x] #7 N+1 audit complete: handle_params, stream rendering, and LiveComponent Repo calls checked
- [x] #8 Findings report written as a Backlog.md document with file:line references, severity ratings, and fix recommendations
<!-- AC:END -->
## Final Summary
<!-- SECTION:FINAL_SUMMARY:BEGIN -->
## Audit Complete: 9 findings across 4 severity levels
**SQL traces captured** via QueryReporter at `/tmp/queries-ml-audit.sql` for all 4 key pages (StatsLive, CollectionLive.Index, CollectionLive.Show, ArtistLive.Show).
### Key findings:
**HIGH severity (3)**:
1. **Display toggle triggers full DB reload** — grid/list toggle in `handle_set_display` re-executes FTS search when CSS-only toggle suffices. Fix: remove `load_and_assign_records` call.
2. **StatsLive.Index mount: 10 sync queries** — home page TTFB blocked by 10 sequential queries. `assign_async` already used elsewhere on the same page — extend it to cover heavy queries.
3. **Scrobble updates trigger full stream reset** — every 5-minute cron batch causes `reset: true` on all connected clients for both StatsLive and ScrobbledTracksLive.
**MEDIUM severity (4)**: 4. **Redundant stream_insert + immediate full reload** in ScrobbleRulesLive and OnlineStoreTemplateLive — `stream_insert` is immediately nullified by `load_and_assign_*`. 5. **Full resets on pagination**`reset: true` on page changes when append could suffice. 6. **Missing optimistic UI** for delete operations — DB waits block UI. 7. **Correlated subqueries** in `ListeningStats.tracks_with_record_info_query` — 3 subqueries per row × 100 tracks.
**PASS (2)**: 8. All search inputs have `phx-debounce` set. 9. No direct N+1 issues found in LiveView/LiveComponent layer (no direct Repo calls).
Full report: `docs/backlog/documents/doc-27` — includes file:line references, SQL traces, and fix recommendations for each finding.
<!-- SECTION:FINAL_SUMMARY:END -->