docs(cache): trim verbose comments to concise intent + KEYS/ARGV specs

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
Jake BarnbyandClaude Opus 4.8 committed 2026-07-01 20:25:53 +12:00
1 parent e834c54eea
commit 38c0b57e72
2 files changed
+27 -88

No files matched your search

+20 -65
View File
@@ -23,32 +23,16 @@ class Redis implements Adapter, Leasable, Retryable
private const GENERATION_FIELD = '__utopia_gen__';
/**
* Reserved hash field holding the epoch-microsecond deadline until which
* saveWithLease() is refused after a purge. The generation lease alone cannot
* reject a reader that captured the post-purge generation yet read the row
* from a source lagging behind the write (a replica, or an older MVCC
* snapshot on a pooled connection): its token matches, so the stale value
* would be cached. This tombstone shrinks that window by refusing every lease
* save for graceWindow ms after a purge, regardless of token. It is a bounded
* mitigation, not a proof: a reader whose read stays stale for longer than
* graceWindow can still slip through, so size graceWindow to the worst-case
* read-after-write staleness (replica lag / snapshot lifetime). Fully closing
* the race needs a database freshness token (LSN/GTID), not a wall clock.
* Lives in the same hash to keep operations single-key. Not a usable field.
* Reserved hash field: epoch-µs deadline until which saveWithLease() is
* refused after a purge, so a reader whose DB read lagged the write can't
* re-cache stale data under a still-valid generation token. Not a cache field.
*/
private const TOMBSTONE_FIELD = '__utopia_tomb__';
/**
* Save $hash field into hash $key only when the key's generation field still
* equals the caller's token AND, when a grace window is set, the key is past
* its post-purge tombstone. redis.call('TIME') is CLOCK_REALTIME, so a
* deadline more than the window ahead of now means the wall clock stepped
* backward since the stamp; it is ignored rather than wedging saves, which
* also bounds any block to graceWindow of real time. A non-numeric window or
* deadline is treated as absent (fail open), and a spent/invalid deadline is
* dropped so a field-purged hash doesn't keep the tombstone forever.
* Single-key. KEYS[1]=key; ARGV[1]=hash, ARGV[2]=value, ARGV[3]=expected
* generation, ARGV[4]=grace window in ms (0 disables it).
* Save $hash only if the generation token matches and the key is past its
* post-purge tombstone. Single-key. KEYS[1]=key; ARGV[1]=hash, ARGV[2]=value,
* ARGV[3]=expected generation, ARGV[4]=grace window ms (0 = tombstone off).
*/
private const LUA_SAVE_WITH_LEASE = <<<'LUA'
local current = redis.call('HGET', KEYS[1], '__utopia_gen__')
@@ -62,6 +46,7 @@ class Redis implements Adapter, Leasable, Retryable
if deadline ~= nil then
local t = redis.call('TIME')
local now = tonumber(t[1]) * 1000000 + tonumber(t[2])
-- 2nd clause ignores a deadline left by a since-rewound clock
if now < deadline and (deadline - now) <= window * 1000 then
return 0
end
@@ -74,16 +59,10 @@ class Redis implements Adapter, Leasable, Retryable
LUA;
/**
* Drop $key's value fields and advance its generation in one step, so an
* in-flight reader holding the previous generation cannot re-cache stale
* data. The bump is unconditional: a reader may have read a row the writer is
* now deleting before anything was cached, so the generation must move even
* when nothing was cached. When ARGV[1] (grace window, ms) is > 0 the key is
* also stamped with a tombstone that refuses lease saves for that long, so a
* reader whose read lags the write cannot repopulate stale data even with a
* matching token. Returns the number of value fields removed (the reserved
* generation and tombstone fields are excluded), preserving purge()
* semantics. Single-key. KEYS[1]=key; ARGV[1]=grace window in ms.
* Drop $key's value fields, advance its generation (so an in-flight reader
* can't re-cache stale data), and stamp the tombstone when ARGV[1] > 0.
* Returns value fields removed (reserved fields excluded), preserving purge()
* semantics. Single-key. KEYS[1]=key; ARGV[1]=grace window ms.
*/
private const LUA_PURGE_BUMP = <<<'LUA'
local gen = '__utopia_gen__'
@@ -103,14 +82,9 @@ class Redis implements Adapter, Leasable, Retryable
LUA;
/**
* Delete a single $hash field and advance the key's generation in one step,
* so a field-level purge invalidates in-flight leases just like a full purge,
* and stamp the post-purge tombstone when a grace window is set. The tombstone
* is key-level, so it also briefly blocks lease re-caching of the key's other
* fields (conservative: never caches stale, may skip a fresh sibling save
* within the window). Returns the number of fields removed, preserving
* purge()'s deletion-result semantics. Single-key. KEYS[1]=key; ARGV[1]=field,
* ARGV[2]=grace window in ms.
* Delete one $hash field, advance the generation, and stamp the tombstone when
* ARGV[2] > 0 (key-level, so it also briefly blocks re-caching sibling fields).
* Single-key. KEYS[1]=key; ARGV[1]=field, ARGV[2]=grace window ms.
*/
private const LUA_PURGE_FIELD = <<<'LUA'
local removed = redis.call('HDEL', KEYS[1], ARGV[1])
@@ -136,12 +110,8 @@ class Redis implements Adapter, Leasable, Retryable
private int $retryDelay = 1000; // milliseconds
/**
* Milliseconds after a purge during which saveWithLease() is refused, to
* defend against a reader whose read lags the purge's write (replica lag or a
* stale MVCC snapshot on a pooled connection). Size it to the worst-case
* read-after-write staleness horizon: a read that stays stale longer than
* this can still slip through. 0 disables the tombstone and preserves the
* pure generation-lease behaviour.
* Milliseconds after a purge during which saveWithLease() is refused. Size to
* the worst-case read-after-write staleness; 0 disables the tombstone.
*/
private int $leaseGraceWindow = 0;
@@ -215,9 +185,7 @@ class Redis implements Adapter, Leasable, Retryable
}
/**
* @param int $milliseconds window after a purge during which lease saves are
* refused; size it to the worst-case read-after-write
* staleness horizon (0 disables the tombstone)
* @param int $milliseconds grace window after a purge (0 disables the tombstone)
* @return self
*/
public function setLeaseGraceWindow(int $milliseconds): self
@@ -232,12 +200,7 @@ class Redis implements Adapter, Leasable, Retryable
return $this->leaseGraceWindow;
}
/**
* The generation and tombstone live in each key's own hash; a caller must
* never read, write or delete them through the public field API, or they
* could reset the generation and revive stale lease tokens, or tamper with
* the post-purge tombstone.
*/
/** Reserved internal fields must never be reachable through the public field API. */
private function isReserved(string $hash): bool
{
return $hash === self::GENERATION_FIELD || $hash === self::TOMBSTONE_FIELD;
@@ -398,10 +361,6 @@ class Redis implements Adapter, Leasable, Retryable
));
}
// Drop the value fields and advance the in-hash generation in one atomic
// step (single-key Redis Lua) so an in-flight reader holding the previous
// generation cannot re-cache stale data after this purge, and stamp the
// post-purge tombstone when a grace window is configured.
return (bool) $this->execute(fn () => $this->redis->eval(
self::LUA_PURGE_BUMP,
[$key, $this->leaseGraceWindow],
@@ -438,12 +397,8 @@ class Redis implements Adapter, Leasable, Retryable
*/
public function getSize(): int
{
// Trade-off of the in-hash generation: a purged key keeps a tiny hash
// holding its reserved generation (and, when a grace window is set,
// tombstone) field until it is re-cached, so it still counts here.
// Filtering those out would need an O(N) scan with a per-key HLEN check;
// for a diagnostic counter that isn't worth it, so DBSIZE may slightly
// over-count purged-but-not-yet-recached keys.
// A purged key keeps its reserved field(s) until re-cached, so DBSIZE can
// slightly over-count purged-but-not-yet-recached keys.
/** @var int $size */
$size = $this->execute(fn () => $this->redis->dbSize());
+7 -23
View File
@@ -160,11 +160,8 @@ class LeasableTest extends TestCase
}
/**
* The tombstone closes the residual gap the generation lease cannot: a reader
* that captured the CURRENT (post-purge) generation but whose read lagged the
* write (a replica, or a stale MVCC snapshot on a pooled connection) would
* otherwise re-cache stale data, because its token matches. Within the
* post-purge grace window, saveWithLease must refuse even a token-valid save.
* A token-valid save (generation captured post-purge) must still be refused
* within the grace window, then allowed once the deadline passes.
*/
public function testTombstoneRejectsTokenValidSaveWithinGraceWindow(): void
{
@@ -175,11 +172,8 @@ class LeasableTest extends TestCase
$cache = new Cache($adapter);
$cache->flush();
// A writer commits and purges; the tombstone opens.
$cache->purge('doc:1');
// The reader's token is valid (captured post-purge) but its read lagged
// the write, so the value it wants to cache is stale.
$generation = $cache->getGeneration('doc:1');
$this->assertFalse(
$cache->saveWithLease('doc:1', ['stale' => true], 'doc:1', $generation),
@@ -187,9 +181,7 @@ class LeasableTest extends TestCase
);
$this->assertFalse($cache->load('doc:1', 60, 'doc:1'), 'Cache must not hold the stale value');
// Simulate the window elapsing by moving the deadline into the past (a
// direct write, no sleep): lease saves resume, and the spent tombstone is
// dropped rather than lingering in the hash.
// Expire the tombstone directly (no sleep) to resume saves.
$past = ((int) \time() - 60) * 1000000;
$redis->hSet('doc:1', '__utopia_tomb__', (string) $past);
@@ -204,10 +196,7 @@ class LeasableTest extends TestCase
);
}
/**
* A field-level purge must open the same tombstone as a full purge, so a
* lagging reader cannot repopulate a stale hash field either.
*/
/** A field-level purge must open the tombstone too. */
public function testFieldPurgeAlsoOpensTombstone(): void
{
$cache = $this->graceCache(500);
@@ -232,8 +221,6 @@ class LeasableTest extends TestCase
$cache->saveWithLease('doc:1', ['v' => 1], 'field-a', $cache->getGeneration('doc:1'));
$cache->purge('doc:1');
// The purged key holds only its reserved generation and tombstone
// fields; neither may surface as a listable cache field.
$this->assertSame([], $cache->list('doc:1'));
}
@@ -249,10 +236,8 @@ class LeasableTest extends TestCase
}
/**
* TIME is CLOCK_REALTIME, so a backward wall-clock step could leave a stamped
* deadline far in the future and wedge every lease save until real time
* caught up. A deadline more than the window ahead of now must be treated as
* a clock anomaly and ignored, not honoured.
* A deadline more than the window ahead of now (a since-rewound clock) must
* be ignored, not wedge saves.
*/
public function testTombstoneIgnoresImplausibleFutureDeadline(): void
{
@@ -266,8 +251,7 @@ class LeasableTest extends TestCase
$cache->purge('doc:1');
$generation = $cache->getGeneration('doc:1');
// A deadline 1h ahead (in microseconds) — far beyond the 500ms window —
// stands in for a backward clock step after the stamp.
// 1h-ahead deadline (µs), far beyond the window: a since-rewound clock.
$farFutureMicros = ((int) \time() + 3600) * 1000000;
$redis->hSet('doc:1', '__utopia_tomb__', (string) $farFutureMicros);