xspeed / includes / class-hit-counter.php

class-hit-counter.php in xSpeed Cache: AI-Powered Performance Hub with MCP, Caching & CDN 1.1.1, at includes/class-hit-counter.php

453 lines 16.1 KB

Raw Download Zip

1	<?php
2	/**
3	* Hit_Counter — rolling 24h hits + misses for cache requests.
4	*
5	* Storage: one transient `xspeed_hit_buffer` containing a list of up to
6	* 24 hourly buckets. Each bucket: [hour_start_ts, hits, misses]. Bucket
7	* keyed by floor(time()/3600); old buckets drop off when we push a new
8	* hour. Transient TTL set to 25 hours so an idle site doesn't lose its
9	* history immediately after going quiet.
10	*
11	* Writes happen on every cached HIT and every MISS (Cache.php records
12	* via the static record_* methods). We absorb the cost in an in-process
13	* static accumulator that flushes to the transient once per request via
14	* register_shutdown_function, so the served-from-disk hot path pays
15	* nothing.
16	*
17	* @package XSpeed
18	*/
19
20	declare(strict_types=1);
21
22	namespace XSpeed;
23
24	defined( 'ABSPATH' ) \|\| exit;
25
26	final class Hit_Counter {
27
28	public const TRANSIENT_KEY = 'xspeed_hit_buffer';
29	public const TTL = 90000; // 25h
30	public const MAX_BUCKETS = 24;
31
32	/**
33	* Option key holding the bucket buffer.
34	*
35	* Why an OPTION, not a transient (fixed 2026-06-16): with a persistent
36	* object cache absent or misconfigured, `set_transient()` writes to the
37	* object cache ONLY (never the DB) when an external object cache is
38	* "in use" — even if that cache is non-persistent (e.g. xSpeed's own
39	* object-cache drop-in falling back to an in-request array because Redis
40	* isn't reachable). In that state every recorded hit/miss was written to
41	* a per-request cache and discarded at request end, so the dashboard
42	* hit-ratio read 0 (or a meaningless 100% off one drained log line).
43	* Options always persist to wp_options, so the counter survives across
44	* requests regardless of the object-cache backend. read_buffer() also
45	* busts the options-group cache entry before reading so a stale
46	* in-request copy from a non-persistent cache can't shadow the DB value.
47	*/
48	public const OPT_KEY = 'xspeed_hit_buffer';
49
50	/** Daily hit/miss aggregates (option, autoload off): 'Y-m-d' => {hits,misses}. */
51	public const DAILY_OPT = 'xspeed_hit_daily';
52
53	/** Days of daily history to retain (the trend UI reads 7/30). */
54	public const DAILY_MAX_DAYS = 120;
55
56	/**
57	* @var array<int,int> Pending increments keyed by metric ('hit'\|'miss').
58	* Flushed to the transient on shutdown.
59	*/
60	private static $pending = array( 'hit' => 0, 'miss' => 0 );
61
62	/**
63	* @var bool Whether the shutdown flush is already registered.
64	*/
65	private static $shutdown_registered = false;
66
67	public static function record_hit(): void {
68	++self::$pending['hit'];
69	self::ensure_shutdown_flush();
70	}
71
72	public static function record_miss(): void {
73	++self::$pending['miss'];
74	// Flush misses INLINE, not at shutdown. A MISS is recorded ONLY here
75	// (HITs additionally have the durable hits.log drain as a backstop),
76	// so if a miss flush is ever dropped the dashboard ratio skews toward
77	// 100%. Flushing inline guarantees the miss is committed to the
78	// options-backed buffer (see OPT_KEY) within this request, before any
79	// shutdown-time object-cache teardown could interfere. Misses are
80	// low-frequency (one per page per cache fill), so the inline write
81	// cost is negligible; HITs stay deferred (high-volume).
82	self::flush_pending();
83	}
84
85	/**
86	* Add `$count` HITs in one shot. Used by collect_nginx_log_hits()
87	* to attribute many HITs served directly by nginx (bypassing PHP)
88	* to the counter once we've drained the log file.
89	*/
90	public static function record_hits_batch( int $count ): void {
91	if ( $count <= 0 ) {
92	return;
93	}
94	self::$pending['hit'] += $count;
95	self::ensure_shutdown_flush();
96	}
97
98	/**
99	* Drain the HITs log file at wp-content/cache/xspeed/hits.log. Two
100	* serve paths that can't call record_hit() inline append one line per
101	* HIT here: the nginx server-level rewrite block (see
102	* Cache::nginx_snippet(), serves without ever reaching PHP) and the
103	* advanced-cache.php drop-in (runs before WordPress loads, so
104	* Hit_Counter isn't available). This method reads the line count,
105	* truncates the file, and folds the count into Hit_Counter via
106	* record_hits_batch — so both uncountable-inline paths still show up
107	* in the dashboard hit-ratio on the next load.
108	*
109	* Returns the number of HITs collected (0 if the log is missing,
110	* empty, or the rewrite block isn't engaged).
111	*
112	* Concurrency: file is opened with LOCK_EX before the read/truncate
113	* round-trip so a concurrent nginx write can't lose entries. Nginx
114	* uses buffer=16k flush=10s on its access_log so writes are batched
115	* and the lock contention is negligible.
116	*/
117	public static function collect_nginx_log_hits(): int {
118	// Lives under uploads/, not the cache dir — see Cache::hits_log_dir()
119	// (FBS-82478: a cache-dir access_log can take nginx down on purge/
120	// uninstall).
121	$path = Cache::hits_log_path();
122	if ( ! file_exists( $path ) ) {
123	return 0;
124	}
125	if ( filesize( $path ) === 0 ) {
126	return 0;
127	}
128	// phpcs:ignore WordPress.WP.AlternativeFunctions.file_system_operations_fopen, WordPress.PHP.NoSilencedErrors.Discouraged -- WP_Filesystem doesn't model fopen+flock+ftruncate atomically; we need the lock to prevent nginx writes from being lost.
129	$fp = @fopen( $path, 'r+' );
130	if ( ! $fp ) {
131	return 0;
132	}
133	// Non-blocking exclusive lock — if nginx is mid-write we just skip
134	// this collection and try again on the next dashboard load.
135	// phpcs:ignore WordPress.WP.AlternativeFunctions.file_system_operations_flock -- See fopen rationale.
136	if ( ! @flock( $fp, LOCK_EX \| LOCK_NB ) ) { // phpcs:ignore WordPress.PHP.NoSilencedErrors.Discouraged
137	fclose( $fp ); // phpcs:ignore WordPress.WP.AlternativeFunctions.file_system_operations_fclose -- pairs with the flock'd fopen above; WP_Filesystem can't model flock.
138	return 0;
139	}
140	$count = 0;
141	while ( ( $line = fgets( $fp ) ) !== false ) {
142	if ( '' !== rtrim( $line ) ) {
143	++$count;
144	}
145	}
146	// phpcs:ignore WordPress.WP.AlternativeFunctions.file_system_operations_ftruncate -- See fopen rationale.
147	ftruncate( $fp, 0 );
148	flock( $fp, LOCK_UN );
149	fclose( $fp ); // phpcs:ignore WordPress.WP.AlternativeFunctions.file_system_operations_fclose -- pairs with the flock'd fopen above; WP_Filesystem can't model flock.
150
151	if ( $count > 0 ) {
152	self::record_hits_batch( $count );
153	// Flush immediately — the next read of totals_24h() happens
154	// inline in Cache::get_stats(), before register_shutdown_function
155	// could fire. Without this, the dashboard sees stale numbers
156	// and the just-drained HITs appear on the FOLLOWING refresh.
157	self::flush_pending();
158	}
159	return $count;
160	}
161
162	/** Option key storing the last-scanned byte offset of the access log. */
163	public const SERVER_LOG_OFFSET_OPT = 'xspeed_access_log_offset';
164
165	/**
166	* Count Apache/LiteSpeed static-rewrite HITs by scanning the web
167	* server's access log.
168	*
169	* On Apache/LiteSpeed a cache HIT is served straight from the
170	* `xspeed-static/` tree by a `.htaccess` RewriteRule — the request
171	* never reaches PHP, so (unlike the nginx path, which logs to our own
172	* dedicated hits.log) there's no inline hook to call record_hit().
173	* Instead we read the server's own access log incrementally: every
174	* request whose logged path contains our static-cache dir was a HIT
175	* served below PHP.
176	*
177	* Incremental + safe:
178	* - We remember a byte offset (SERVER_LOG_OFFSET_OPT) and only read
179	* bytes appended since last time — O(new traffic), not O(log size).
180	* - If the log shrank (rotation/truncation) we reset the offset to 0
181	* and rescan from the top once, so a rotation never double-counts
182	* or permanently desyncs.
183	* - We never write to the log, only read; failure is silent.
184	*
185	* Returns 0 (and is a no-op) when no readable access log exists — the
186	* common managed-host case. The drop-in/PHP path still counts its own
187	* HITs, so hit-ratio degrades to "PHP-served hits only" rather than 0.
188	*
189	* @return int HITs folded in this call.
190	*/
191	public static function collect_server_log_hits(): int {
192	// Apache only. nginx writes its own dedicated hits.log (drained by
193	// collect_nginx_log_hits); LiteSpeed routes hits through the PHP
194	// drop-in (which also appends to that hits.log) because its
195	// .htaccess can't header/log a static serve — see
196	// Cache::static_rewrite_allowed(). So Apache is the lone server that
197	// serves static hits below PHP yet logs them to the SERVER's access
198	// log, which is what we scan here.
199	if ( Server::APACHE !== Server::type() ) {
200	return 0;
201	}
202
203	$path = Server::access_log_path();
204	if ( '' === $path ) {
205	return 0;
206	}
207
208	$size = @filesize( $path ); // phpcs:ignore WordPress.PHP.NoSilencedErrors.Discouraged -- log may vanish on rotation between checks.
209	if ( false === $size ) {
210	return 0;
211	}
212
213	$offset = (int) get_option( self::SERVER_LOG_OFFSET_OPT, 0 );
214	if ( $offset > $size ) {
215	// Log was rotated/truncated since last scan — start over so we
216	// don't seek past EOF and miss the new file's lines.
217	$offset = 0;
218	}
219	if ( $offset === $size ) {
220	return 0; // Nothing new since last drain.
221	}
222
223	// phpcs:ignore WordPress.WP.AlternativeFunctions.file_system_operations_fopen, WordPress.PHP.NoSilencedErrors.Discouraged -- read-only incremental tail of an external log; WP_Filesystem can't fseek and would buffer the whole file through memory.
224	$fp = @fopen( $path, 'r' );
225	if ( ! $fp ) {
226	return 0;
227	}
228	if ( $offset > 0 ) {
229	fseek( $fp, $offset );
230	}
231
232	// The static-cache dir, as it appears in a logged request path. We
233	// match on the request-target substring so the access-log format
234	// (combined/common/custom) doesn't matter — every format includes
235	// the request line.
236	$needle = '/' . trim( str_replace( ABSPATH, '', XSPEED_CACHE_STATIC_DIR ), '/' );
237	$count = 0;
238	while ( ( $line = fgets( $fp ) ) !== false ) {
239	// Only count GET requests that landed on the static tree. The
240	// "GET " + needle pairing avoids counting our own loopback
241	// probe writes or unrelated dir listings.
242	if ( false !== strpos( $line, $needle ) && false !== strpos( $line, 'GET ' ) ) {
243	++$count;
244	}
245	}
246	$new_offset = ftell( $fp );
247	fclose( $fp ); // phpcs:ignore WordPress.WP.AlternativeFunctions.file_system_operations_fclose -- pairs with the read-only fopen above.
248
249	// Persist the offset even when count is 0 so we don't re-scan the
250	// same non-matching bytes every dashboard load.
251	update_option( self::SERVER_LOG_OFFSET_OPT, (int) $new_offset, false );
252
253	if ( $count > 0 ) {
254	self::record_hits_batch( $count );
255	self::flush_pending();
256	}
257	return $count;
258	}
259
260	/**
261	* Returns up to MAX_BUCKETS most-recent hourly buckets oldest →
262	* newest. Each bucket: [ts => unix hour-start, hits => int, misses
263	* => int ].
264	*
265	* @return array<int,array{ts:int,hits:int,misses:int}>
266	*/
267	/**
268	* Read the bucket buffer straight from the options table, busting any
269	* stale per-request object-cache copy first so a non-persistent cache
270	* can never shadow the committed DB value. See OPT_KEY docblock.
271	*
272	* @return mixed Raw stored value (array on success).
273	*/
274	private static function read_buffer() {
275	// Drop the cached 'options' entry for our key so get_option() falls
276	// through to the DB. Harmless on a persistent cache (it just reloads
277	// from the DB once); essential on a non-persistent one.
278	\wp_cache_delete( self::OPT_KEY, 'options' );
279	return get_option( self::OPT_KEY, array() );
280	}
281
282	private static function write_buffer( array $buf ): void {
283	// Autoload 'no' — the buffer is read only in admin/stats contexts, so
284	// it must never inflate the frontend alloptions payload.
285	if ( false === get_option( self::OPT_KEY, false ) ) {
286	add_option( self::OPT_KEY, $buf, '', 'no' );
287	return;
288	}
289	update_option( self::OPT_KEY, $buf );
290	}
291
292	public static function buckets(): array {
293	$buf = self::read_buffer();
294	if ( ! is_array( $buf ) ) {
295	return array();
296	}
297	// Defensive — strip anything not shaped right.
298	$out = array();
299	foreach ( $buf as $b ) {
300	if ( is_array( $b ) && isset( $b['ts'], $b['hits'], $b['misses'] ) ) {
301	$out[] = array(
302	'ts' => (int) $b['ts'],
303	'hits' => (int) $b['hits'],
304	'misses' => (int) $b['misses'],
305	);
306	}
307	}
308	return $out;
309	}
310
311	/**
312	* Totals over the last 24h (sum across all buckets).
313	*
314	* @return array{hits:int,misses:int,ratio:float}
315	*/
316	public static function totals_24h(): array {
317	$buckets = self::buckets();
318	$hits = 0;
319	$misses = 0;
320	foreach ( $buckets as $b ) {
321	$hits += $b['hits'];
322	$misses += $b['misses'];
323	}
324	$total = $hits + $misses;
325	return array(
326	'hits' => $hits,
327	'misses' => $misses,
328	'ratio' => $total > 0 ? round( $hits / $total, 4 ) : 0.0,
329	);
330	}
331
332	public static function reset(): void {
333	delete_transient( self::TRANSIENT_KEY );
334	// The bucket buffer lives in the OPT_KEY option (migrated off the
335	// transient); reset() must clear it too, or record→reset leaves the
336	// old hit/miss buckets behind and buckets() still reports them.
337	delete_option( self::OPT_KEY );
338	\wp_cache_delete( self::OPT_KEY, 'options' );
339	delete_option( self::SERVER_LOG_OFFSET_OPT );
340	delete_option( self::DAILY_OPT );
341	self::$pending = array( 'hit' => 0, 'miss' => 0 );
342	}
343
344	/**
345	* One-shot register on first record_* call this request.
346	*/
347	private static function ensure_shutdown_flush(): void {
348	if ( self::$shutdown_registered ) {
349	return;
350	}
351	self::$shutdown_registered = true;
352	register_shutdown_function( array( __CLASS__, 'flush_pending' ) );
353	}
354
355	/**
356	* Flush in-process counters into the transient. Bucketed by current
357	* hour. New hour → append a bucket and drop the oldest if we exceed
358	* MAX_BUCKETS.
359	*/
360	public static function flush_pending(): void {
361	$pending = self::$pending;
362	if ( 0 === $pending['hit'] && 0 === $pending['miss'] ) {
363	return;
364	}
365	self::$pending = array( 'hit' => 0, 'miss' => 0 );
366
367	$hour = (int) ( time() - ( time() % 3600 ) );
368	$buf = self::buckets();
369	$last = end( $buf );
370	$updated = false;
371
372	if ( $last && $last['ts'] === $hour ) {
373	$buf[ count( $buf ) - 1 ]['hits'] += $pending['hit'];
374	$buf[ count( $buf ) - 1 ]['misses'] += $pending['miss'];
375	$updated = true;
376	}
377
378	if ( ! $updated ) {
379	$buf[] = array(
380	'ts' => $hour,
381	'hits' => $pending['hit'],
382	'misses' => $pending['miss'],
383	);
384	while ( count( $buf ) > self::MAX_BUCKETS ) {
385	array_shift( $buf );
386	}
387	}
388
389	self::write_buffer( $buf );
390	self::bump_daily( $pending['hit'], $pending['miss'] );
391	}
392
393	/**
394	* Fold the just-flushed counts into the persistent daily series. The
395	* hourly buckets expire after ~25h; this option is what makes 7/30-day
396	* hit-ratio trends possible (issue #44). Autoload off — it's only read
397	* by the dashboard/REST, never on the frontend hot path.
398	*/
399	private static function bump_daily( int $hits, int $misses ): void {
400	if ( $hits <= 0 && $misses <= 0 ) {
401	return;
402	}
403	$day = gmdate( 'Y-m-d' );
404	$series = get_option( self::DAILY_OPT, array() );
405	if ( ! is_array( $series ) ) {
406	$series = array();
407	}
408	if ( ! isset( $series[ $day ] ) \|\| ! is_array( $series[ $day ] ) ) {
409	$series[ $day ] = array(
410	'hits' => 0,
411	'misses' => 0,
412	);
413	}
414	$series[ $day ]['hits'] += $hits;
415	$series[ $day ]['misses'] += $misses;
416	if ( count( $series ) > self::DAILY_MAX_DAYS ) {
417	ksort( $series );
418	$series = array_slice( $series, -self::DAILY_MAX_DAYS, null, true );
419	}
420	update_option( self::DAILY_OPT, $series, false );
421	}
422
423	/**
424	* The stored daily hit/miss series, oldest→newest, at most $days rows.
425	*
426	* @return array<int,array{date:string,hits:int,misses:int,ratio:float}>
427	*/
428	public static function daily_series( int $days = 30 ): array {
429	$series = get_option( self::DAILY_OPT, array() );
430	if ( ! is_array( $series ) \|\| empty( $series ) ) {
431	return array();
432	}
433	ksort( $series );
434	$series = array_slice( $series, -max( 1, $days ), null, true );
435	$out = array();
436	foreach ( $series as $date => $row ) {
437	if ( ! is_array( $row ) ) {
438	continue;
439	}
440	$hits = (int) ( $row['hits'] ?? 0 );
441	$misses = (int) ( $row['misses'] ?? 0 );
442	$total = $hits + $misses;
443	$out[] = array(
444	'date' => (string) $date,
445	'hits' => $hits,
446	'misses' => $misses,
447	'ratio' => $total > 0 ? round( $hits / $total, 4 ) : 0.0,
448	);
449	}
450	return $out;
451	}
452	}
453