/**
* Purge a key from all clusters
*
- * This instantiates a hold-off period where the key cannot be
- * written to avoid race conditions where dependent keys get updated
- * with a stale value (e.g. from a DB slave). This is implemented by
- * storing a special "tombstone" value at the cache key that this
- * class recognizes; get() calls will return false for the key.
+ * This deletes the key and instantiates a hold-off period where the key
+ * cannot be written to for the next few seconds (HOLDOFF_TTL). This is to
+ * avoid the following race condition:
+ * a) Some DB data changes and delete() is called on a corresponding key
+ * b) A request refills the key with a stale value from a lagged DB
+ * c) The stale value is stuck there until the key is expired/evicted
+ *
+ * This is implemented by storing a special "tombstone" value at the cache
+ * key that this class recognizes; get() calls will return false for the key
+ * and any set() calls will refuse to replace tombstone values at the key.
+ * For this to always avoid writing stale values, the following must hold:
+ * a) Replication lag is bounded to being less than HOLDOFF_TTL; or
+ * b) If lag is higher, the DB will have gone into read-only mode already
*
* This should only be called when the underlying data (being cached)
* changes in a significant way. If called twice on the same key, then