Re: [PATCH v4 4/6] revision: implement sparse algorithm

"Derrick Stolee via GitGitGadget" [off-list ref] writes:

From: Derrick Stolee <redacted>

When enumerating objects to place in a pack-file during 'git
pack-objects --revs', we discover the "frontier" of commits
that we care about and the boundary with commit we find
uninteresting. From that point, we walk trees to discover which
trees and blobs are uninteresting. Finally, we walk trees from the
interesting commits to find the interesting objects that are
placed in the pack.

This commit introduces a new, "sparse" way to discover the
uninteresting trees. We use the perspective of a single user trying
to push their topic to a large repository. That user likely changed
a very small fraction of the paths in their working directory, but
we spend a lot of time walking all reachable trees.

The way to switch the logic to work in this sparse way is to start
caring about which paths introduce new trees. While it is not
possible to generate a diff between the frontier boundary and all
of the interesting commits, we can simulate that behavior by
inspecting all of the root trees as a whole, then recursing down
to the set of trees at each path.

We already had taken the first step by passing an oidset to
mark_trees_uninteresting_sparse(). We now create a dictionary
whose keys are paths and values are oidsets. We consider the set
of trees that appear at each path. While we inspect a tree, we
add its subtrees to the oidsets corresponding to the tree entry's
path. We also mark trees as UNINTERESTING if the tree we are
parsing is UNINTERESTING.

To actually improve the peformance, we need to terminate our

performance?

recursion. If the oidset contains only UNINTERESTING trees, then

So, at this point, it is not quite clear what "the oidset" refers to
to me.  When you are looking at the root tree of one commit, you'd
add the object ID of the tree to the dictionary keyed by the path
(i.e. root of the tree).  You traverse to the parent and keep adding
these root trees, and after a while you'd hit a negative commit and
add its tree with UNINTERESTING bit.  Before that point, the oidset
for the root of the tree has only interesting trees, so you won't
recurse into its contents, but now at the final point you'd walk to
propagate UNINTERESTING bit down?

we do not continue the recursion. This avoids walking trees that
are likely to not be reachable from interesting trees. If the
oidset contains only interesting trees, then we will walk these
trees in the final stage that collects the intersting objects to
place in the pack. Thus, we only recurse if the oidset contains
both interesting and UNINITERESTING trees.

There are a few ways that this is not a universally better option.

First, we can pack extra objects. If someone copies a subtree
from one tree to another, the first tree will appear UNINTERESTING
and we will not recurse to see that the subtree should also be
UNINTERESTING. We will walk the new tree and see the subtree as
a "new" object and add it to the pack. We add a test case that
demonstrates this as a way to prove that the --sparse option is
actually working.

That's an interesting use of the word "working" ;-)

Second, we can have extra memory pressure. If instead of being a
single user pushing a small topic we are a server sending new
objects from across the entire working directory, then we will
gain very little (the recursion will rarely terminate early) but
will spend extra time maintaining the path-oidset dictionaries.
...
I used the number of walked trees the main metric above because
it is consistent across multiple runs. When I ran my tests, the
performance of the pack-objects command with the same options
could change the end-to-end time by 10x depending on the file
system being warm. However, by repeating the same test on repeat
I could get more consistent timing results. The git.git and
Linux tests were too fast overall (less than 0.5s) to measure
an end-to-end difference. The Azure DevOps case was slow enough
to see the time improve from 15s to 1s in the warm case. The
cold case was 90s to 9s in my testing.

Understandable.  Projects that tend to be deep (e.g. Java?) would
likely to benefit if they are structured and partitioned very well.

+struct path_and_oids_entry {
+	struct hashmap_entry ent;
+	char *path;
+	struct oidset set;

Again "set"?  If I understand the description of the logic explained
in the proposed commit log message correctly, then this is a set of
"tree objects at this path", right?  At least calling this "trees"
may give us slightly better indication.

+static int path_and_oids_cmp(const void *hashmap_cmp_fn_data,
+			     const struct path_and_oids_entry *e1,
+			     const struct path_and_oids_entry *e2,
+			     const void *keydata)
+{
+	return strcmp(e1->path, e2->path);
+}
+
+int map_flags = 0;

That's too bland a name for a system-wide global variable.  If it is
a file-scope "static", that might be more acceptable.

In any case, let BSS take care of the initialization and do not
initialize an int explicitly to 0.

Is it even used?  At least at this step the comparison callback
function does not seem to make any use of the fn-data, so a better
organization may be to delay the introduction of this variable to a
later step where it actually gets used, at which time, the code may
have a better context to give the variable a more suitable name.

+static void paths_and_oids_init(struct hashmap *map)
+{
+	hashmap_init(map, (hashmap_cmp_fn) path_and_oids_cmp, &map_flags, 0);
+}
+
+static void paths_and_oids_clear(struct hashmap *map)
+{
+	struct hashmap_iter iter;
+	struct path_and_oids_entry *entry;
+	hashmap_iter_init(map, &iter);
+
+	while ((entry = (struct path_and_oids_entry *)hashmap_iter_next(&iter))) {
+		oidset_clear(&entry->set);
+		free(entry->path);
+	}
+
+	hashmap_free(map, 1);
+}
+
+static void paths_and_oids_insert(struct hashmap *map,
+				  const char *path,
+				  const struct object_id *oid)
+{
+	int hash = strhash(path);
+	struct path_and_oids_entry key;
+	struct path_and_oids_entry *entry;
+
+	hashmap_entry_init(&key, hash);
+	key.path = xstrdup(path);
+	oidset_init(&key.set, 0);
+
+	if (!(entry = (struct path_and_oids_entry *)hashmap_get(map, &key, NULL))) {
+		entry = xcalloc(1, sizeof(struct path_and_oids_entry));
+		hashmap_entry_init(entry, hash);
+		entry->path = key.path;
+		oidset_init(&entry->set, 16);
+		hashmap_put(map, entry);
+	} else {
+		free(key.path);
+	}

Hmph, having to always allocate only to free sounds wasteful.  

Can't we instead do the look-up with "path" that are given read-only
to us, and allocate only when we are creating the entry in the
dictionary for the first time?  After all, we'd make many look-ups
of the same path to add trees for the same path and only one of them
(i.e. the first one) is used to create an entry in the dictionary.

Or does it not matter?

+
+	oidset_insert(&entry->set, oid);
+}
+
+static void add_children_by_path(struct repository *r,
+				 struct tree *tree,
+				 struct hashmap *map)
+{
+	struct tree_desc desc;
+	struct name_entry entry;
+
+	if (!tree)
+		return;
+
+	if (parse_tree_gently(tree, 1) < 0)
+		return;
+
+	init_tree_desc(&desc, tree->buffer, tree->size);
+	while (tree_entry(&desc, &entry)) {
+		switch (object_type(entry.mode)) {
+		case OBJ_TREE:
+			paths_and_oids_insert(map, entry.path, entry.oid);
+
+			if (tree->object.flags & UNINTERESTING) {
+				struct tree *child = lookup_tree(r, entry.oid);
+				if (child)
+					child->object.flags |= UNINTERESTING;
+			}
+			break;
+		case OBJ_BLOB:
+			if (tree->object.flags & UNINTERESTING) {
+				struct blob *child = lookup_blob(r, entry.oid);
+				if (child)
+					child->object.flags |= UNINTERESTING;
+			}
+			break;
+		default:
+			/* Subproject commit - not in this repository */
+			break;
+		}
+	}
+
+	free_tree_buffer(tree);
+}

OK.

 void mark_trees_uninteresting_sparse(struct repository *r,
 				     struct oidset *set)
 {
+	unsigned has_interesting = 0, has_uninteresting = 0;
+	struct hashmap map;
+	struct hashmap_iter map_iter;
+	struct path_and_oids_entry *entry;
 	struct object_id *oid;
 	struct oidset_iter iter;
 
 	oidset_iter_init(set, &iter);
-	while ((oid = oidset_iter_next(&iter))) {
+	while ((!has_interesting || !has_uninteresting) &&
+	       (oid = oidset_iter_next(&iter))) {
 		struct tree *tree = lookup_tree(r, oid);
 
 		if (!tree)
 			continue;
 
+		if (tree->object.flags & UNINTERESTING)
+			has_uninteresting = 1;
+		else
+			has_interesting = 1;
+	}
+
+	/* Do not walk unless we have both types of trees. */
+	if (!has_uninteresting || !has_interesting)
+		return;

OK.

OK, the change above sort of makes sense ;-)

And the reason why f4 and its contents are wastefully sent is
because the whole tree is duplicated, which is understandable.

 	git pack-objects --stdout --revs --sparse <packinput.txt >sparse.pack &&
 	git index-pack -o sparse.idx sparse.pack &&
 	git show-index <sparse.idx | awk "{print \$2}" >sparse_objects.txt &&
-	test_cmp expect_objects.txt sparse_objects.txt
+	test_cmp expect_sparse_objects.txt sparse_objects.txt
 '
 
 test_done