A custom module's hook runs on every request and the site slows down

hook_preprocess_html() counting unpublished nodes. This runs on every page render, including anonymous pages that should be fully cached. A COUNT(*) over 30,000 nodes is cheap in isolation, but it defeats the render cache: any output that depends on a query result must either declare a cache tag or be uncacheable. Because the preprocess doesn't declare tags, the page can't be safely cached with that value, so the query runs per request. That's the 200ms → 1.8s regression for anonymous users.

hook_entity_presave() loading 40,000 terms. This runs on every entity save - not just the content type that needs the check. Loading all 40,000 terms into memory to do a name match is O(n) per save and pulls the entire vocabulary into PHP on each of the 200 editors' saves. That's the 5-8s save time. It also runs for config entities, users, and taxonomy terms themselves, multiplying the cost.

The developer's max-age = 0 suggestion makes it worse: it explicitly marks the output uncacheable, guaranteeing the query runs on every request forever.

Preprocess html: don't query in the preprocess. Either:

• Compute the count in a block/controller that declares #cache['tags'] = ['node_list'] so it invalidates when nodes change, or
• If it must be in the html preprocess, attach the value via a lazy builder or a #attached render array with #cache['tags'] = ['node_list'] and a sensible max-age. Never max-age = 0.

The count only changes on node save, so node_list is the correct tag - it invalidates exactly when needed and lets the page cache serve anonymous users between saves.

Presave: scope it and bound it.

function acme_utils_entity_presave(EntityInterface $entity) {
  if ($entity->getEntityTypeId() !== 'node' || $entity->bundle() !== 'article') {
    return;
  }
  $name = $entity->label();
  $ids = \Drupal::entityQuery('taxonomy_term')
    ->accessCheck(FALSE)
    ->condition('vid', 'topics')
    ->condition('name', $name)
    ->range(0, 1)
    ->execute();
  if ($ids) {
    $entity->set('field_topic', reset($ids));
  }
}

This does one indexed lookup instead of loading 40,000 terms, and only for the bundle that needs it. If the vocabulary has no index on name, add one - the entity query will use it.

Two tests, both cheap:

1. A kernel test for the presave hook. Create an article node, save it, assert the term reference is set; create a node of a different bundle, save it, and assert the hook did nothing (this catches the missing bundle check). Assert the number of queries executed stays constant as the vocabulary grows - or simply assert the hook doesn't call loadMultiple on all terms, which is the actual defect.

2. A functional test for the html preprocess. Request a page as anonymous, assert the response is cacheable (the X-Drupal-Cache header or the render array's #cache metadata contains node_list and a non-zero max-age). Then save a node and assert the page reflects the new count - this catches both the missing tag and the max-age = 0 regression.

Neither test needs a browser. Both would have failed on the original code and pass on the fix. The point is to assert cacheability metadata, because the performance bug was invisible in the rendered output.