+ private static final Logger LOG = LoggerFactory.getLogger(YangStoreServiceImpl.class);
+
+ /**
+ * This is a rather interesting locking model. We need to guard against both the
+ * cache expiring from GC and being invalidated by schema context change. The
+ * context can change while we are doing processing, so we do not want to block
+ * it, so no synchronization can happen on the methods.
+ *
+ * So what we are doing is the following:
+ *
+ * We synchronize with GC as usual, using a SoftReference.
+ *
+ * The atomic reference is used to synchronize with {@link #refresh()}, e.g. when
+ * refresh happens, it will push a SoftReference(null), e.g. simulate the GC. Now
+ * that may happen while the getter is already busy acting on the old schema context,
+ * so it needs to understand that a refresh has happened and retry. To do that, it
+ * attempts a CAS operation -- if it fails, in knows that the SoftReference has
+ * been replaced and thus it needs to retry.
+ *
+ * Note that {@link #getYangStoreSnapshot()} will still use synchronize() internally
+ * to stop multiple threads doing the same work.
+ */
+ private final AtomicReference<SoftReference<YangStoreSnapshotImpl>> ref = new AtomicReference<>(new SoftReference<YangStoreSnapshotImpl>(null));