# Tableau prepare_table_data Cache Implementation Plan
> **For agentic workers:** REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (`- [ ]`) syntax for tracking.
**Goal:** Make page navigation, sort changes, and repeated filter visits in the `/tableau` page near-instant by memoizing the expensive filter+sort+post-process pipeline inside `prepare_table_data`.
**Architecture:** Extract a memoized inner function `_load_filter_sort_postprocess(filter_query, sort_by_key)` that performs the heavy work (load full data, filter, sort, collect, cast-to-string, fill-null, add HTML links, format values) and returns a fully post-processed Polars DataFrame. The outer `prepare_table_data` becomes a thin wrapper that handles non-deterministic side effects (`track_search`, `uuid.uuid4()` for cleanup trigger, `data_timestamp + 1`) and pagination. The memoized helper only runs when no `data` argument is passed (i.e., the Tableau path). Other callers (`acheteur`, `titulaire`, `observatoire`) keep the current uncached path because they pass an externally-provided LazyFrame that is not safely hashable for cache keys.
**Tech Stack:** Polars (LazyFrame, DataFrame), Flask-Caching (`@cache.memoize()` on `FileSystemCache` already configured in `src/app.py:38`), pytest for unit tests.
**Git**: the issue id is #72, add the reference in commit messages.
---
## Background and constraints
Read these before starting; they explain why the design takes the shape it does.
1. **Cache infrastructure is already wired.** `src/cache.py` defines `cache = Cache()`. `src/app.py:38-48` initializes it with `FileSystemCache`, default 24h timeout, `CACHE_THRESHOLD=300`. The cache directory is wiped on every restart (`rmtree` at `src/app.py:36`), so cache always starts empty.
2. **Existing pattern to mirror.** `src/pages/observatoire.py:650-660` already uses `@cache.memoize()` plus a `_normalize_filter_params` helper that converts a dict of filters into a hashable tuple. This plan applies the same idiom to `sort_by` (which is a `list[dict]` from Dash DataTable).
3. **Non-deterministic outputs that MUST stay outside the memoized function:**
- `data_timestamp + 1` (increments each call; would freeze if cached)
- `trigger_cleanup = str(uuid.uuid4())` (intentionally unique per call to fire the clientside filter-cleanup callback)
- `track_search(filter_query, source_table)` — Matomo HTTP POST, currently called inside `filter_table_data` at `src/utils/table.py:214`. Must fire on every user action including cache hits.
4. **Tracking call site move.** `track_search` must move OUT of `filter_table_data` and into each caller, otherwise cache hits would silently skip Matomo tracking. Current callers of `filter_table_data` to update:
- `src/utils/table.py:402` (inside `prepare_table_data`)
- `src/pages/tableau.py:325` (`download_data` callback)
- `src/pages/acheteur.py:427` (`download_data_acheteur` callback)
- `src/pages/titulaire.py:443` (`download_data_titulaire` callback)
5. **Why Tableau-only caching.** `prepare_table_data` is also called from `acheteur.py`, `titulaire.py`, `observatoire.py`. Those callers pass a pre-filtered LazyFrame or list-of-dicts as `data`. Hashing arbitrary LazyFrames or large lists for memoization is impractical. The fix gates on `data is None` (the Tableau path) and leaves the other paths byte-for-byte identical.
6. **Cache key composition.** The memoized function takes only `(filter_query, sort_by_key)`. `page_current` and `page_size` are intentionally NOT in the key — pagination happens in the outer wrapper after retrieving the cached, fully post-processed frame. This means every page click and page-size change is a cache hit (the whole point of the change).
7. **Pickling.** Flask-Caching pickles arguments to form keys and pickles return values to disk. Polars `DataFrame` pickles cleanly. `LazyFrame` does not — so the memoized function must `.collect()` before returning.
8. **File path expectations.** All paths below are relative to repo root `/home/colin/git/decp.info`. Run all commands from there.
---
## File Structure
- **Modify** `src/utils/table.py` — extract memoized helper, refactor `prepare_table_data`, remove `track_search` call from `filter_table_data`.
- **Modify** `src/pages/tableau.py` — add explicit `track_search` call in `download_data`.
- **Modify** `src/pages/acheteur.py` — add explicit `track_search` call in `download_data_acheteur`.
- **Modify** `src/pages/titulaire.py` — add explicit `track_search` call in `download_data_titulaire`.
- **Create** `tests/test_table.py` — unit tests for new helpers and refactored `prepare_table_data`.
---
## Task 1: Set up unit tests for table.py
**Files:**
- Create: `tests/test_table.py`
This task scaffolds a non-Selenium pytest module so subsequent tasks can do TDD without booting a Dash server. The conftest already writes a small `tests/test.parquet` fixture (see `tests/conftest.py:10`); reuse it.
- [ ] **Step 1: Write the failing test**
Create `tests/test_table.py` with:
```python
import os
import polars as pl
import pytest
@pytest.fixture
def sample_lff():
"""Small LazyFrame with the columns needed by add_links / format_values."""
return pl.LazyFrame(
[
{
"uid": "u1",
"id": "u1",
"acheteur_id": "12345678900011",
"acheteur_nom": "Mairie de Test",
"titulaire_id": "98765432100022",
"titulaire_nom": "Entreprise Test",
"titulaire_typeIdentifiant": "SIRET",
"objet": "Travaux divers",
"montant": 12500.0,
"dateNotification": "2025-03-15",
"codeCPV": "45000000",
"dureeRestanteMois": 6,
"titulaire_distance": 42.0,
}
]
)
def test_table_module_imports():
from src.utils import table
assert hasattr(table, "prepare_table_data")
```
- [ ] **Step 2: Run test to verify it passes (sanity check)**
Run: `uv run pytest tests/test_table.py -v`
Expected: PASS for `test_table_module_imports`. (Selenium is not invoked because no `dash_duo` fixture is used.)
- [ ] **Step 3: Commit**
```bash
git add tests/test_table.py
git commit -m "test: scaffold unit tests for table utilities"
```
---
## Task 2: Move track_search out of filter_table_data
**Files:**
- Modify: `src/utils/table.py:210-274` (remove `track_search` import usage at line 214)
- Modify: `src/pages/tableau.py:317-334` (`download_data` callback)
- Modify: `src/pages/acheteur.py:425-430` area (`download_data_acheteur` callback)
- Modify: `src/pages/titulaire.py:441-446` area (`download_data_titulaire` callback)
- Modify: `tests/test_table.py` (add a test that confirms `filter_table_data` no longer calls Matomo)
`track_search` must move out so that the soon-to-be-memoized helper does not swallow tracking on cache hits. We do this BEFORE introducing caching so that the diff is small and verifiable on its own.
- [ ] **Step 1: Write the failing test**
Append to `tests/test_table.py`:
```python
def test_filter_table_data_does_not_call_track_search(monkeypatch, sample_lff):
from src.utils import table
calls = []
monkeypatch.setattr(table, "track_search", lambda *a, **kw: calls.append(a))
result = table.filter_table_data(
sample_lff, "{objet} icontains travaux", "tableau"
).collect()
assert calls == []
assert result.height == 1
```
- [ ] **Step 2: Run test to verify it fails**
Run: `uv run pytest tests/test_table.py::test_filter_table_data_does_not_call_track_search -v`
Expected: FAIL (`assert calls == []` fails because `filter_table_data` currently calls `track_search` at line 214).
- [ ] **Step 3: Remove the track_search call from filter_table_data**
Edit `src/utils/table.py` — find this block:
```python
def filter_table_data(
lff: pl.LazyFrame, filter_query: str, filter_source: str
) -> pl.LazyFrame:
_schema = lff.collect_schema()
track_search(filter_query, filter_source)
filtering_expressions = filter_query.split(" && ")
```
Remove the `track_search(filter_query, filter_source)` line. Result:
```python
def filter_table_data(
lff: pl.LazyFrame, filter_query: str, filter_source: str
) -> pl.LazyFrame:
_schema = lff.collect_schema()
filtering_expressions = filter_query.split(" && ")
```
The `filter_source` parameter remains in the signature (avoids changing all callers in this task). It becomes unused; that is acceptable since callers will pass it again later if needed. Do NOT remove the `from src.utils.tracking import track_search` import yet — `prepare_table_data` will use it in Task 5.
- [ ] **Step 4: Add explicit track_search calls in download callbacks**
In `src/pages/tableau.py`, find:
```python
def download_data(n_clicks, filter_query, sort_by, hidden_columns: list = None):
lff: pl.LazyFrame = query_marches().lazy()
# Les colonnes masquées sont supprimées
if hidden_columns:
lff = lff.drop(hidden_columns)
if filter_query:
lff = filter_table_data(lff, filter_query, "tab download")
```
Insert a `track_search` call so behavior is preserved. First add the import at the top of `src/pages/tableau.py` next to other `src.utils` imports:
```python
from src.utils.tracking import track_search
```
Then change the body:
```python
def download_data(n_clicks, filter_query, sort_by, hidden_columns: list = None):
lff: pl.LazyFrame = query_marches().lazy()
# Les colonnes masquées sont supprimées
if hidden_columns:
lff = lff.drop(hidden_columns)
if filter_query:
track_search(filter_query, "tab download")
lff = filter_table_data(lff, filter_query, "tab download")
```
Repeat the same pattern in `src/pages/acheteur.py` (search for `filter_table_data(lff, filter_query, "ach download")`):
Add import:
```python
from src.utils.tracking import track_search
```
Wrap the call:
```python
if filter_query:
track_search(filter_query, "ach download")
lff = filter_table_data(lff, filter_query, "ach download")
```
Repeat in `src/pages/titulaire.py` (search for `filter_table_data(lff, filter_query, "titu download")`):
Add import:
```python
from src.utils.tracking import track_search
```
Wrap the call:
```python
if filter_query:
track_search(filter_query, "titu download")
lff = filter_table_data(lff, filter_query, "titu download")
```
- [ ] **Step 5: Run test to verify it passes**
Run: `uv run pytest tests/test_table.py::test_filter_table_data_does_not_call_track_search -v`
Expected: PASS.
- [ ] **Step 6: Run full unit test file to verify no regressions**
Run: `uv run pytest tests/test_table.py -v`
Expected: All tests in `test_table.py` PASS.
- [ ] **Step 7: Commit**
```bash
git add src/utils/table.py src/pages/tableau.py src/pages/acheteur.py src/pages/titulaire.py tests/test_table.py
git commit -m "refactor: move track_search out of filter_table_data into callers"
```
---
## Task 3: Add normalize_sort_by helper
**Files:**
- Modify: `src/utils/table.py` (add helper near other utility functions, e.g. after `dates_to_strings`)
- Modify: `tests/test_table.py` (add tests)
A cache key must be hashable. Dash DataTable's `sort_by` is a `list[dict]` like `[{"column_id": "montant", "direction": "asc"}, ...]`, which is not hashable. We mirror the `_normalize_filter_params` idiom from `src/pages/observatoire.py:650-657`.
- [ ] **Step 1: Write the failing tests**
Append to `tests/test_table.py`:
```python
def test_normalize_sort_by_handles_empty():
from src.utils.table import normalize_sort_by
assert normalize_sort_by(None) == ()
assert normalize_sort_by([]) == ()
def test_normalize_sort_by_returns_hashable_tuple():
from src.utils.table import normalize_sort_by
sort_by = [
{"column_id": "montant", "direction": "desc"},
{"column_id": "dateNotification", "direction": "asc"},
]
key = normalize_sort_by(sort_by)
assert key == (("montant", "desc"), ("dateNotification", "asc"))
# Must be hashable so that flask-caching can build a cache key from it
hash(key)
def test_normalize_sort_by_preserves_order():
"""Order matters for sort: [A, B] != [B, A]."""
from src.utils.table import normalize_sort_by
a_then_b = normalize_sort_by(
[{"column_id": "a", "direction": "asc"}, {"column_id": "b", "direction": "asc"}]
)
b_then_a = normalize_sort_by(
[{"column_id": "b", "direction": "asc"}, {"column_id": "a", "direction": "asc"}]
)
assert a_then_b != b_then_a
```
- [ ] **Step 2: Run tests to verify they fail**
Run: `uv run pytest tests/test_table.py -v -k normalize_sort_by`
Expected: FAIL with `ImportError` for `normalize_sort_by`.
- [ ] **Step 3: Implement normalize_sort_by**
Edit `src/utils/table.py`. Add this function immediately after the `dates_to_strings` function (around line 148):
```python
def normalize_sort_by(sort_by) -> tuple:
"""Convert Dash DataTable sort_by (list[dict]) into a hashable tuple
suitable for use as a cache key. Order is preserved because it determines
sort precedence."""
if not sort_by:
return ()
return tuple((entry["column_id"], entry["direction"]) for entry in sort_by)
```
- [ ] **Step 4: Run tests to verify they pass**
Run: `uv run pytest tests/test_table.py -v -k normalize_sort_by`
Expected: 3 PASS.
- [ ] **Step 5: Commit**
```bash
git add src/utils/table.py tests/test_table.py
git commit -m "feat: add normalize_sort_by hashable cache-key helper"
```
---
## Task 4: Extract memoized post-process helper
**Files:**
- Modify: `src/utils/table.py` (add `_load_filter_sort_postprocess`, decorate with `@cache.memoize()`, import `cache`)
- Modify: `tests/test_table.py` (add tests)
Introduce the function whose result will live in the FileSystemCache. Inputs: `(filter_query, sort_by_key)`. Output: a fully post-processed, unpaginated Polars DataFrame ready to slice and convert to dicts.
This task does NOT yet wire the helper into `prepare_table_data` — that happens in Task 5. Splitting these tasks keeps each diff small and testable.
- [ ] **Step 1: Write the failing tests**
Append to `tests/test_table.py`:
```python
@pytest.fixture(autouse=True)
def reset_cache():
"""Ensure the flask-caching backend is empty between tests so that
cache-hit assertions are meaningful. Falls back to no-op when no
Flask app context is active (NullCache)."""
from utils.cache import cache
try:
cache.clear()
except RuntimeError:
# No app context — cache is NullCache, nothing to clear
pass
yield
def test_load_filter_sort_postprocess_returns_dataframe(monkeypatch, sample_lff):
from src.utils import table
monkeypatch.setattr(
table, "query_marches", lambda: sample_lff.collect()
)
df = table._load_filter_sort_postprocess(filter_query=None, sort_by_key=())
assert isinstance(df, pl.DataFrame)
assert df.height == 1
# All values must be strings after post-processing
for col in df.columns:
assert df.schema[col] == pl.String
def test_load_filter_sort_postprocess_applies_filter(monkeypatch, sample_lff):
from src.utils import table
monkeypatch.setattr(
table, "query_marches", lambda: sample_lff.collect()
)
df = table._load_filter_sort_postprocess(
filter_query="{objet} icontains travaux", sort_by_key=()
)
assert df.height == 1
df_empty = table._load_filter_sort_postprocess(
filter_query="{objet} icontains nonexistent", sort_by_key=()
)
assert df_empty.height == 0
def test_load_filter_sort_postprocess_adds_links(monkeypatch, sample_lff):
from src.utils import table
monkeypatch.setattr(
table, "query_marches", lambda: sample_lff.collect()
)
df = table._load_filter_sort_postprocess(filter_query=None, sort_by_key=())
# add_links injects an wrapper around uid, acheteur_nom, titulaire_nom
assert " 0:
dff = format_values(dff)
return dff
```
- [ ] **Step 4: Run tests to verify they pass**
Run: `uv run pytest tests/test_table.py -v -k load_filter_sort_postprocess`
Expected: 3 PASS.
- [ ] **Step 5: Run the full test_table.py to catch regressions**
Run: `uv run pytest tests/test_table.py -v`
Expected: All PASS.
- [ ] **Step 6: Commit**
```bash
git add src/utils/table.py tests/test_table.py
git commit -m "feat: add memoized _load_filter_sort_postprocess helper"
```
---
## Task 5: Wire the memoized helper into prepare_table_data
**Files:**
- Modify: `src/utils/table.py` — replace the body of `prepare_table_data` so the Tableau path uses the cache
- Modify: `tests/test_table.py` — add tests covering the new flow
The outer function keeps its signature unchanged so callers in `acheteur.py`, `titulaire.py`, `observatoire.py`, `tableau.py` need no updates. When `data is None` (the Tableau case), use the memoized helper; otherwise fall through to the original logic.
- [ ] **Step 1: Write the failing tests**
Append to `tests/test_table.py`:
```python
def test_prepare_table_data_returns_expected_tuple(monkeypatch, sample_lff):
from src.utils import table
monkeypatch.setattr(
table, "query_marches", lambda: sample_lff.collect()
)
result = table.prepare_table_data(
data=None,
data_timestamp=5,
filter_query=None,
page_current=0,
page_size=20,
sort_by=[],
source_table="tableau",
)
# Same arity as before: 9 outputs
assert len(result) == 9
dicts, columns, tooltip, ts, nb_rows, dl_disabled, dl_text, dl_title, cleanup = result
assert isinstance(dicts, list)
assert ts == 6 # data_timestamp + 1 must still increment
assert "1 lignes" in nb_rows
def test_prepare_table_data_calls_track_search_on_filter(monkeypatch, sample_lff):
from src.utils import table
calls = []
monkeypatch.setattr(
table, "query_marches", lambda: sample_lff.collect()
)
monkeypatch.setattr(table, "track_search", lambda *a, **kw: calls.append(a))
table.prepare_table_data(
data=None,
data_timestamp=0,
filter_query="{objet} icontains travaux",
page_current=0,
page_size=20,
sort_by=[],
source_table="tableau",
)
assert calls == [("{objet} icontains travaux", "tableau")]
def test_prepare_table_data_paginates_without_recomputing(monkeypatch, sample_lff):
"""Two calls with same filter+sort but different pages must invoke
the inner heavy work only once."""
from src.utils import table
call_count = {"n": 0}
real_query = sample_lff.collect()
def counting_query():
call_count["n"] += 1
return real_query
monkeypatch.setattr(table, "query_marches", counting_query)
# First call: cache miss
table.prepare_table_data(
data=None,
data_timestamp=0,
filter_query=None,
page_current=0,
page_size=10,
sort_by=[],
source_table="tableau",
)
first_count = call_count["n"]
# Second call, different page: cache hit, query_marches must NOT fire again
table.prepare_table_data(
data=None,
data_timestamp=0,
filter_query=None,
page_current=1,
page_size=10,
sort_by=[],
source_table="tableau",
)
assert call_count["n"] == first_count, (
"query_marches was called again — pagination triggered cache miss"
)
def test_prepare_table_data_cleanup_trigger_for_non_tableau(monkeypatch, sample_lff):
"""Non-tableau pages still get a fresh uuid trigger, not no_update."""
from dash import no_update
from src.utils import table
monkeypatch.setattr(
table, "query_marches", lambda: sample_lff.collect()
)
result = table.prepare_table_data(
data=None,
data_timestamp=0,
filter_query="{objet} icontains travaux",
page_current=0,
page_size=20,
sort_by=[],
source_table="acheteur",
)
cleanup = result[8]
assert cleanup is not no_update
assert isinstance(cleanup, str)
assert len(cleanup) >= 32 # uuid4 hex string
def test_prepare_table_data_with_external_data_does_not_use_cache(
monkeypatch, sample_lff
):
"""When a caller passes data (acheteur/titulaire/observatoire path),
bypass the memoized helper entirely."""
from src.utils import table
sentinel = {"called": False}
def should_not_be_called(*a, **kw):
sentinel["called"] = True
raise AssertionError("Memoized helper must not be called when data is provided")
monkeypatch.setattr(
table, "_load_filter_sort_postprocess", should_not_be_called
)
table.prepare_table_data(
data=sample_lff, # external LazyFrame
data_timestamp=0,
filter_query=None,
page_current=0,
page_size=20,
sort_by=[],
source_table="acheteur",
)
assert sentinel["called"] is False
```
- [ ] **Step 2: Run tests to verify they fail**
Run: `uv run pytest tests/test_table.py -v -k prepare_table_data`
Expected: At least the cache-hit (`paginates_without_recomputing`) and `track_search`-routing tests FAIL because the current `prepare_table_data` re-runs the full pipeline on every call and routes tracking through `filter_table_data` (which Task 2 already neutralized — so tracking would be lost without the new explicit call).
- [ ] **Step 3: Refactor prepare_table_data**
Edit `src/utils/table.py`. Replace the entire `prepare_table_data` function body with:
```python
def prepare_table_data(
data, data_timestamp, filter_query, page_current, page_size, sort_by, source_table
):
"""
Préparation des données pour les datatables.
Pour la page Tableau (data is None), le calcul lourd (chargement complet,
filtre, tri, post-traitement) est mémorisé via _load_filter_sort_postprocess.
Les changements de page deviennent ainsi des cache hits.
Pour les autres pages (data fourni), le chemin original est conservé : la
LazyFrame externe n'est pas hashable et le coût de filtre/tri y est déjà
minime puisque les données sont pré-restreintes.
"""
logger.debug(" + + + + + + + + + + + + + + + + + + ")
# Side effect non-cacheable : le tracking doit firer sur chaque action
# utilisateur, y compris sur cache hit.
if filter_query:
track_search(filter_query, source_table)
# Trigger uuid pour les pages autres que tableau (clientside cleanup)
trigger_cleanup = (
no_update if source_table == "tableau" else str(uuid.uuid4())
)
if data is None:
# Tableau path : utilise le cache
sort_by_key = normalize_sort_by(sort_by)
dff: pl.DataFrame = _load_filter_sort_postprocess(
filter_query=filter_query, sort_by_key=sort_by_key
)
else:
# acheteur / titulaire / observatoire path : code original, non caché
if isinstance(data, list):
lff: pl.LazyFrame = pl.LazyFrame(
data, strict=False, infer_schema_length=5000
)
elif isinstance(data, pl.LazyFrame):
lff = data
else:
lff = query_marches().lazy()
if filter_query:
lff = filter_table_data(lff, filter_query, source_table)
if sort_by and len(sort_by) > 0:
lff = sort_table_data(lff, sort_by)
dff = lff.collect()
dff = dff.cast(pl.String)
dff = dff.fill_null("")
dff = add_links(dff)
if "sourceFile" in dff.columns:
dff = add_resource_link(dff)
if dff.height > 0:
dff = format_values(dff)
height = dff.height
if height > 0:
nb_rows = (
f"{format_number(height)} lignes "
f"({format_number(dff.select('uid').unique().height)} marchés)"
)
else:
nb_rows = "0 lignes (0 marchés)"
# Pagination — toujours hors cache pour rester sur des cache hits
start_row = page_current * page_size
dff = dff.slice(start_row, page_size)
table_columns, tooltip = setup_table_columns(dff)
dicts = dff.to_dicts()
download_disabled, download_text, download_title = get_button_properties(height)
return (
dicts,
table_columns,
tooltip,
data_timestamp + 1,
nb_rows,
download_disabled,
download_text,
download_title,
trigger_cleanup,
)
```
Notes on what changed vs the original at `src/utils/table.py:372-458`:
- `track_search` now called explicitly at the top, on every invocation (not via `filter_table_data`).
- `data is None` branch delegates the heavy work to the memoized helper.
- `data is not None` branch is functionally identical to the original (pagination still happens after collect+post-process).
- The post-processing (`cast`, `fill_null`, `add_links`, `add_resource_link`, `format_values`) is now done in BOTH branches before `nb_rows` calculation. In the cached branch this was already done inside `_load_filter_sort_postprocess`; in the uncached branch we keep doing it inline. This means `nb_rows` and `dff.select('uid').unique().height` operate on the post-processed frame in both branches, matching the original semantics.
- [ ] **Step 4: Run all unit tests**
Run: `uv run pytest tests/test_table.py -v`
Expected: All PASS, including `test_prepare_table_data_paginates_without_recomputing`.
- [ ] **Step 5: Run the full repo test suite to catch regressions**
Run: `uv run pytest -v`
Expected: All PASS. Selenium tests (`tests/test_main.py`) require Chrome/Chromium; if the executor lacks a browser, those tests will error/skip — note the failures and rerun in an environment with Chrome before declaring done.
- [ ] **Step 6: Commit**
```bash
git add src/utils/table.py tests/test_table.py
git commit -m "perf(tableau): memoize filter+sort+postprocess pipeline"
```
---
## Task 6: Manual smoke test in the browser
**Files:** none modified.
Type checks and unit tests cannot validate that page navigation actually feels faster. This task is explicitly a hands-on verification.
- [ ] **Step 1: Start the dev server**
Run: `uv run run.py`
Wait for `Dash is running on http://...`.
- [ ] **Step 2: Open the Tableau page and warm the cache**
1. Open `http://localhost:8050/tableau` (or whatever port the dev server prints).
2. With no filter applied, wait for the first page to load fully. This is the cold-cache load (slow expected).
3. Open the browser devtools Network panel.
- [ ] **Step 3: Verify pagination is fast**
1. Click "page 2" / "page 3" / "page 4" in the table footer in quick succession.
2. Each navigation should return data in well under 1 second (in the original code each took several seconds).
3. In the dev server logs, look for the line `Cache miss — recomputing for filter=...` from `_load_filter_sort_postprocess`. It should appear ONCE for the initial load and NOT appear again as you change pages.
- [ ] **Step 4: Verify a new filter triggers exactly one cache miss**
1. In the table, type a filter into one of the columns (e.g. `paris` in `acheteur_commune_nom`) and press Enter.
2. The dev log should show ONE new `Cache miss — recomputing` line.
3. Change page within the filtered view — no new cache miss line should appear.
- [ ] **Step 5: Verify filter cleanup trigger still fires**
1. Open `http://localhost:8050/acheteur?id=` (use any valid id from the dataset).
2. Apply a filter on the embedded table.
3. The clientside callback for filter cleanup (`src/assets/dash_clientside.js` `clean_filters`) should still rewrite the filter operators (e.g. `contains` → `icontains`). If it doesn't fire, the `trigger_cleanup` uuid is broken — investigate.
- [ ] **Step 6: Verify download still works**
1. On the Tableau page, click "Télécharger au format Excel" (the button must be enabled — apply a filter that brings the row count under 65,000).
2. The downloaded XLSX must open and contain the filtered rows.
- [ ] **Step 7: Stop the dev server**
Ctrl-C.
- [ ] **Step 8: If all checks pass, this completes the implementation**
No commit — this task is verification only. Report results to the user.