named_caches: make more resilient to corrupted state

client/named_cache.py

Index: client/named_cache.py
diff --git a/client/named_cache.py b/client/named_cache.py
index 50a4e1c04fb9005f5d4a83150f3ee405f0044e63..e62062595a2951cfee5551a2115a5b4db665b204 100644
--- a/client/named_cache.py
+++ b/client/named_cache.py
@@ -63,12 +63,17 @@ class CacheManager(object):
 
     Returns a context manager that must be closed as soon as possible.
     """
-    state_path = os.path.join(self.root_dir, u'state.json')
     with self._lock:
+      state_path = os.path.join(self.root_dir, u'state.json')
+      assert self._lru is None, 'acquired lock, but self._lru is not None'
       if os.path.isfile(state_path):
-        self._lru = lru.LRUDict.load(state_path)
-      else:
-        self._lru = lru.LRUDict()
+        try:
+          self._lru = lru.LRUDict.load(state_path)
+        except ValueError as ex:
+          logging.exception('failed to load named cache state file')
+          logging.warning('deleting named caches')
+          file_path.rmtree(self.root_dir)

[M-A Ruel, 2017/03/17 13:43:41]

nit: I'd prefer the directory itself to not be deleted but as a quick fix, lgtm

[nodir, 2017/03/17 16:20:56]

why? CacheManager works fine if the directory doesn't exist.

[M-A Ruel, 2017/03/20 23:22:57]

I don't know, it's a feeling about simplifying edge cases, but it doesn't
matter. It was optional.

+      self._lru = self._lru or lru.LRUDict()
       if time_fn:
         self._lru.time_fn = time_fn
       try: