Merge pull request #1535 from LLNL/bugfix/faster-install-db-gh1521

[WIP] Faster database loading, faster in-memory hashing

lib/spack/spack/database.py

@@ -164,6 +164,9 @@ def __init__(self, root, db_dir=None):
         self.lock = Lock(self._lock_path)
         self._data = {}
 
+        # whether there was an error at the start of a read transaction
+        self._error = None
+
     def write_transaction(self, timeout=_db_lock_timeout):
         """Get a write lock context manager for use in a `with` block."""
         return WriteTransaction(self.lock, self._read, self._write, timeout)
@@ -198,7 +201,7 @@ def _write_to_yaml(self, stream):
         except YAMLError as e:
             raise SpackYAMLError("error writing YAML database:", str(e))
 
-    def _read_spec_from_yaml(self, hash_key, installs, parent_key=None):
+    def _read_spec_from_yaml(self, hash_key, installs):
         """Recursively construct a spec from a hash in a YAML database.
 
         Does not do any locking.
@@ -212,19 +215,27 @@ def _read_spec_from_yaml(self, hash_key, installs, parent_key=None):
 
         # Build spec from dict first.
         spec = Spec.from_node_dict(spec_dict)
+        return spec
 
+    def _assign_dependencies(self, hash_key, installs, data):
         # Add dependencies from other records in the install DB to
         # form a full spec.
+        spec = data[hash_key].spec
+        spec_dict = installs[hash_key]['spec']
+
         if 'dependencies' in spec_dict[spec.name]:
             yaml_deps = spec_dict[spec.name]['dependencies']
             for dname, dhash, dtypes in Spec.read_yaml_dep_specs(yaml_deps):
-                child = self._read_spec_from_yaml(dhash, installs, hash_key)
-                spec._add_dependency(child, dtypes)
+                if dhash not in data:
+                    tty.warn("Missing dependency not in database: ",
+                             "%s needs %s-%s" % (
+                                 spec.format('$_$#'), dname, dhash[:7]))
+                    continue
 
-        # Specs from the database need to be marked concrete because
-        # they represent actual installations.
-        spec._mark_concrete()
-        return spec
+                # defensive copy (not sure everything handles extra
+                # parent links yet)
+                child = data[dhash].spec
+                spec._add_dependency(child, dtypes)
 
     def _read_from_yaml(self, stream):
         """
@@ -248,7 +259,8 @@ def _read_from_yaml(self, stream):
 
         def check(cond, msg):
             if not cond:
-                raise CorruptDatabaseError(self._index_path, msg)
+                raise CorruptDatabaseError(
+                    "Spack database is corrupt: %s" % msg, self._index_path)
 
         check('database' in yfile, "No 'database' attribute in YAML.")
 
@@ -267,22 +279,28 @@ def check(cond, msg):
             self.reindex(spack.install_layout)
             installs = dict((k, v.to_dict()) for k, v in self._data.items())
 
-        # Iterate through database and check each record.
+        def invalid_record(hash_key, error):
+            msg = ("Invalid record in Spack database: "
+                   "hash: %s, cause: %s: %s")
+            msg %= (hash_key, type(e).__name__, str(e))
+            raise CorruptDatabaseError(msg, self._index_path)
+
+        # Build up the database in three passes:
+        #
+        #   1. Read in all specs without dependencies.
+        #   2. Hook dependencies up among specs.
+        #   3. Mark all specs concrete.
+        #
+        # The database is built up so that ALL specs in it share nodes
+        # (i.e., its specs are a true Merkle DAG, unlike most specs.)
+
+        # Pass 1: Iterate through database and build specs w/o dependencies
         data = {}
         for hash_key, rec in installs.items():
             try:
                 # This constructs a spec DAG from the list of all installs
                 spec = self._read_spec_from_yaml(hash_key, installs)
 
-                # Validate the spec by ensuring the stored and actual
-                # hashes are the same.
-                spec_hash = spec.dag_hash()
-                if not spec_hash == hash_key:
-                    tty.warn(
-                        "Hash mismatch in database: %s -> spec with hash %s" %
-                        (hash_key, spec_hash))
-                    continue  # TODO: is skipping the right thing to do?
-
                 # Insert the brand new spec in the database.  Each
                 # spec has its own copies of its dependency specs.
                 # TODO: would a more immmutable spec implementation simplify
@@ -290,11 +308,22 @@ def check(cond, msg):
                 data[hash_key] = InstallRecord.from_dict(spec, rec)
 
             except Exception as e:
-                tty.warn("Invalid database reecord:",
-                         "file:  %s" % self._index_path,
-                         "hash:  %s" % hash_key,
-                         "cause: %s: %s" % (type(e).__name__, str(e)))
-                raise
+                invalid_record(hash_key, e)
+
+        # Pass 2: Assign dependencies once all specs are created.
+        for hash_key in data:
+            try:
+                self._assign_dependencies(hash_key, installs, data)
+            except Exception as e:
+                invalid_record(hash_key, e)
+
+        # Pass 3: Mark all specs concrete.  Specs representing real
+        # installations must be explicitly marked.
+        # We do this *after* all dependencies are connected because if we
+        # do it *while* we're constructing specs,it causes hashes to be
+        # cached prematurely.
+        for hash_key, rec in data.items():
+            rec.spec._mark_concrete()
 
         self._data = data
 
@@ -304,7 +333,26 @@ def reindex(self, directory_layout):
         Locks the DB if it isn't locked already.
 
         """
-        with self.write_transaction():
+        # Special transaction to avoid recursive reindex calls and to
+        # ignore errors if we need to rebuild a corrupt database.
+        def _read_suppress_error():
+            try:
+                if os.path.isfile(self._index_path):
+                    self._read_from_yaml(self._index_path)
+            except CorruptDatabaseError as e:
+                self._error = e
+                self._data = {}
+
+        transaction = WriteTransaction(
+            self.lock, _read_suppress_error, self._write, _db_lock_timeout)
+
+        with transaction:
+            if self._error:
+                tty.warn(
+                    "Spack database was corrupt. Will rebuild. Error was:",
+                    str(self._error))
+                self._error = None
+
             old_data = self._data
             try:
                 self._data = {}
@@ -313,10 +361,15 @@ def reindex(self, directory_layout):
                 for spec in directory_layout.all_specs():
                     # Create a spec for each known package and add it.
                     path = directory_layout.path_for_spec(spec)
-                    old_info = old_data.get(spec.dag_hash())
+
+                    # Try to recover explicit value from old DB, but
+                    # default it to False if DB was corrupt.
                     explicit = False
-                    if old_info is not None:
-                        explicit = old_info.explicit
+                    if old_data is not None:
+                        old_info = old_data.get(spec.dag_hash())
+                        if old_info is not None:
+                            explicit = old_info.explicit
+
                     self._add(spec, path, directory_layout, explicit=explicit)
 
                 self._check_ref_counts()
@@ -601,11 +654,7 @@ def missing(self, spec):
 
 
 class CorruptDatabaseError(SpackError):
-
-    def __init__(self, path, msg=''):
-        super(CorruptDatabaseError, self).__init__(
-            "Spack database is corrupt: %s.  %s." % (path, msg),
-            "Try running `spack reindex` to fix.")
+    """Raised when errors are found while reading the database."""
 
 
 class InvalidDatabaseVersionError(SpackError):

lib/spack/spack/spec.py

@@ -504,6 +504,7 @@ def __init__(self, spec_like, *dep_like, **kwargs):
         self.variants.spec = self
         self.namespace = other.namespace
         self._hash = other._hash
+        self._cmp_key_cache = other._cmp_key_cache
 
         # Specs are by default not assumed to be normal, but in some
         # cases we've read them from a file want to assume normal.
@@ -858,9 +859,10 @@ def return_val(res):
         # Edge traversal yields but skips children of visited nodes
         if not (key in visited and cover == 'edges'):
             # This code determines direction and yields the children/parents
+
             successors = deps
             if direction == 'parents':
-                successors = self.dependents_dict()
+                successors = self.dependents_dict()  # TODO: deptype?
 
             visited.add(key)
             for name in sorted(successors):
@@ -1278,15 +1280,15 @@ def concretize(self):
         # Mark everything in the spec as concrete, as well.
         self._mark_concrete()
 
-    def _mark_concrete(self):
+    def _mark_concrete(self, value=True):
         """Mark this spec and its dependencies as concrete.
 
         Only for internal use -- client code should use "concretize"
         unless there is a need to force a spec to be concrete.
         """
         for s in self.traverse(deptype_query=alldeps):
-            s._normal = True
-            s._concrete = True
+            s._normal = value
+            s._concrete = value
 
     def concretized(self):
         """This is a non-destructive version of concretize().  First clones,
@@ -1533,6 +1535,10 @@ def normalize(self, force=False):
         if self._normal and not force:
             return False
 
+        # avoid any assumptions about concreteness when forced
+        if force:
+            self._mark_concrete(False)
+
         # Ensure first that all packages & compilers in the DAG exist.
         self.validate_names()
         # Get all the dependencies into one DependencyMap
@@ -1865,7 +1871,7 @@ def virtual_dependencies(self):
         """Return list of any virtual deps in this spec."""
         return [spec for spec in self.traverse() if spec.virtual]
 
-    def _dup(self, other, **kwargs):
+    def _dup(self, other, deps=True, cleardeps=True):
         """Copy the spec other into self.  This is an overwriting
            copy.  It does not copy any dependents (parents), but by default
            copies dependencies.
@@ -1896,7 +1902,7 @@ def _dup(self, other, **kwargs):
         self.versions = other.versions.copy()
         self.architecture = other.architecture
         self.compiler = other.compiler.copy() if other.compiler else None
-        if kwargs.get('cleardeps', True):
+        if cleardeps:
             self._dependents = DependencyMap()
             self._dependencies = DependencyMap()
         self.compiler_flags = other.compiler_flags.copy()
@@ -1906,19 +1912,15 @@ def _dup(self, other, **kwargs):
         self.external_module = other.external_module
         self.namespace = other.namespace
         self._hash = other._hash
+        self._cmp_key_cache = other._cmp_key_cache
 
         # If we copy dependencies, preserve DAG structure in the new spec
-        if kwargs.get('deps', True):
+        if deps:
             # This copies the deps from other using _dup(deps=False)
-            # XXX(deptype): We can keep different instances of specs here iff
-            #               it is only a 'build' dependency (from its parent).
-            #               All other instances must be shared (due to symbol
-            #               and PATH contention). These should probably search
-            #               for any existing installation which can satisfy the
-            #               build and latch onto that because if 3 things need
-            #               the same build dependency and it is *not*
-            #               available, we only want to build it once.
-            new_nodes = other.flat_dependencies(deptype_query=alldeps)
+            deptypes = alldeps
+            if isinstance(deps, (tuple, list)):
+                deptypes = deps
+            new_nodes = other.flat_dependencies(deptypes=deptypes)
             new_nodes[self.name] = self
 
             stack = [other]
@@ -1927,6 +1929,9 @@ def _dup(self, other, **kwargs):
                 new_spec = new_nodes[cur_spec.name]
 
                 for depspec in cur_spec._dependencies.values():
+                    if not any(d in deptypes for d in depspec.deptypes):
+                        continue
+
                     stack.append(depspec.spec)
 
                     # XXX(deptype): add any new deptypes that may have appeared
@@ -1942,13 +1947,22 @@ def _dup(self, other, **kwargs):
         self.external_module = other.external_module
         return changed
 
-    def copy(self, **kwargs):
+    def copy(self, deps=True):
         """Return a copy of this spec.
-           By default, returns a deep copy.  Supply dependencies=False
-           to get a shallow copy.
+
+        By default, returns a deep copy. To control how dependencies are
+        copied, supply:
+
+        deps=True:  deep copy
+
+        deps=False: shallow copy (no dependencies)
+
+        deps=('link', 'build'):
+            only build and link dependencies.  Similar for other deptypes.
+
         """
         clone = Spec.__new__(Spec)
-        clone._dup(self, **kwargs)
+        clone._dup(self, deps=deps)
         return clone
 
     @property
@@ -2059,10 +2073,17 @@ def _cmp_key(self):
           1. A tuple describing this node in the DAG.
           2. The hash of each of this node's dependencies' cmp_keys.
         """
-        dep_dict = self.dependencies_dict(deptype=('link', 'run'))
-        return self._cmp_node() + (
-            tuple(hash(dep_dict[name])
-                  for name in sorted(dep_dict)),)
+        if self._cmp_key_cache:
+            return self._cmp_key_cache
+
+        dep_tuple = tuple(
+            (d.spec.name, hash(d.spec), tuple(sorted(d.deptypes)))
+            for name, d in sorted(self._dependencies.items()))
+
+        key = (self._cmp_node(), dep_tuple)
+        if self._concrete:
+            self._cmp_key_cache = key
+        return key
 
     def colorized(self):
         return colorize_spec(self)
@@ -2457,6 +2478,7 @@ def spec(self, name, check_valid_token=False):
         spec._dependencies = DependencyMap()
         spec.namespace = spec_namespace
         spec._hash = None
+        spec._cmp_key_cache = None
 
         spec._normal = False
         spec._concrete = False

lib/spack/spack/test/directory_layout.py

@@ -91,22 +91,33 @@ def test_read_and_write_spec(self):
 
             # Make sure spec file can be read back in to get the original spec
             spec_from_file = self.layout.read_spec(spec_path)
-            self.assertEqual(spec, spec_from_file)
-            self.assertTrue(spec.eq_dag, spec_from_file)
+
+            # currently we don't store build dependency information when
+            # we write out specs to the filesystem.
+
+            # TODO: fix this when we can concretize more loosely based on
+            # TODO: what is installed. We currently omit these to
+            # TODO: increase reuse of build dependencies.
+            stored_deptypes = ('link', 'run')
+            expected = spec.copy(deps=stored_deptypes)
+            self.assertEqual(expected, spec_from_file)
+            self.assertTrue(expected.eq_dag, spec_from_file)
             self.assertTrue(spec_from_file.concrete)
 
             # Ensure that specs that come out "normal" are really normal.
             with open(spec_path) as spec_file:
                 read_separately = Spec.from_yaml(spec_file.read())
 
-                read_separately.normalize()
-                self.assertEqual(read_separately, spec_from_file)
+                # TODO: revise this when build deps are in dag_hash
+                norm = read_separately.normalized().copy(deps=stored_deptypes)
+                self.assertEqual(norm, spec_from_file)
 
-                read_separately.concretize()
-                self.assertEqual(read_separately, spec_from_file)
+                # TODO: revise this when build deps are in dag_hash
+                conc = read_separately.concretized().copy(deps=stored_deptypes)
+                self.assertEqual(conc, spec_from_file)
 
             # Make sure the hash of the read-in spec is the same
-            self.assertEqual(spec.dag_hash(), spec_from_file.dag_hash())
+            self.assertEqual(expected.dag_hash(), spec_from_file.dag_hash())
 
             # Ensure directories are properly removed
             self.layout.remove_install_directory(spec)