[PATCH 0/1] ref-filter.c: faster multi-pattern ref filtering

[PATCH 0/1] ref-filter.c: faster multi-pattern ref filtering

[Taylor Blau]

Hi,

Peff and I have been experimenting with using the references from a
repository's alternate to speed up the connectivity check following a
'git clone --reference'.

We have noticed that the connectivity check becomes much faster when
advertising both the heads and tags of an alternate, as opposed to just
the heads. But, in our initial experiments, we noticed that it took
*far* longer to compute the answer to:

  $ git for-each-ref refs/heads refs/tags

then simply

  $ git for-each-ref refs/heads

We found that this dates back to cfe004a5a9 (ref-filter: limit
traversal to prefix, 2017-05-22), which drops the user-provided patterns
entirely when more than one pattern is passed to the ref filter.

To remedy this, we implemented an algorithm which computes the longest
pattern prefixes over the disjoint subsets of patterns provided to the
ref filter. This produces the most-specific queries (i.e., the ones that
we hope to return the fewest extra results) without covering the same
ref more than once.

The details of the algorithm are written up in detail in the patch. This
doesn't have quite the impressive results on repositories the size of
the kernel, but yields a drastic improvement on our fork network
repositories. Some synthetic results are included in the patch as well.

Thanks in advance for your review.

Taylor Blau (1):
  ref-filter.c: find disjoint pattern prefixes

 ref-filter.c            | 89 +++++++++++++++++++++++++++++------------
 t/t6300-for-each-ref.sh | 26 ++++++++++++
 2 files changed, 89 insertions(+), 26 deletions(-)

--
2.21.0.203.g358da99528

[PATCH 1/1] ref-filter.c: find disjoint pattern prefixes

[Taylor Blau]

Since cfe004a5a9 (ref-filter: limit traversal to prefix, 2017-05-22),
the ref-filter code has sought to limit the traversals to a prefix of
the given patterns.

That code stopped short of handling more than one pattern, because it
means invoking 'for_each_ref_in' multiple times. If we're not careful
about which patterns overlap, we will output the same refs multiple
times.

For instance, consider the set of patterns 'refs/heads/a/*',
'refs/heads/a/b/c', and 'refs/tags/v1.0.0'. If we naïvely ran:

  for_each_ref_in("refs/heads/a/*", ...);
  for_each_ref_in("refs/heads/a/b/c", ...);
  for_each_ref_in("refs/tags/v1.0.0", ...);

we would see 'refs/heads/a/b/c' (and everything underneath it) twice.

Instead, we want to partition the patterns into disjoint sets, where we
know that no ref will be matched by any two patterns in different sets.
In the above, these are:

  - {'refs/heads/a/*', 'refs/heads/a/b/c'}, and
  - {'refs/tags/v1.0.0'}

Given one of these disjoint sets, what is a suitable pattern to pass to
'for_each_ref_in'? One approach is to compute the longest common prefix
over all elements in that disjoint set, and let the caller cull out the
refs they didn't want. Computing the longest prefix means that in most
cases, we won't match too many things the caller would like to ignore.

The longest common prefixes of the above are:

  - {'refs/heads/a/*', 'refs/heads/a/b/c'} -> refs/heads/a/*
  - {'refs/tags/v1.0.0'}                   -> refs/tags/v1.0.0

We instead invoke:

  for_each_ref_in("refs/heads/a/*", ...);
  for_each_ref_in("refs/tags/v1.0.0", ...);

Which provides us with the refs we were looking for with a minimal
amount of extra cruft, but never a duplicate of the ref we asked for.

Implemented here is an algorithm which accomplishes the above, which
works as follows:

  1. Lexicographically sort the given list of patterns.

  2. Initialize 'prefix' to the empty string, where our goal is to
     build each element in the above set of longest common prefixes.

  3. Consider each pattern in the given set, and emit 'prefix' if it
     reaches the end of a pattern, or touches a wildcard character. The
     end of a string is treated as if it precedes a wildcard. (Note that
     there is some room for future work to detect that, e.g., 'a?b' and
     'abc' are disjoint).

  4. Otherwise, recurse on step (3) with the slice of the list
     corresponding to our current prefix (i.e., the subset of patterns
     that have our prefix as a literal string prefix.)

This algorithm is 'O(kn + n log(n))', where 'k' is max(len(pattern)) for
each pattern in the list, and 'n' is len(patterns).

By discovering this set of interesting patterns, we reduce the runtime
of multi-pattern 'git for-each-ref' (and other ref traversals) from
O(N) to O(n log(N)), where 'N' is the total number of packed references.

Running 'git for-each-ref refs/tags/a refs/tags/b' on a repository with
10,000,000 refs in 'refs/tags/huge-N', my best-of-five times drop from:

  real    0m5.805s
  user    0m5.188s
  sys     0m0.468s

to:

  real    0m0.001s
  user    0m0.000s
  sys     0m0.000s

On linux.git, the times to dig out two of the latest -rc tags drops from
0.002s to 0.001s, so the change on repositories with fewer tags is much
less noticeable.

Co-authored-by: Jeff King <peff@peff.net>
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Taylor Blau <me@ttaylorr.com>
---
 ref-filter.c            | 89 +++++++++++++++++++++++++++++------------
 t/t6300-for-each-ref.sh | 26 ++++++++++++
 2 files changed, 89 insertions(+), 26 deletions(-)

diff --git a/ref-filter.c b/ref-filter.c
index 8500671bc6..3e10fd647b 100644
--- a/ref-filter.c
+++ b/ref-filter.c
@@ -20,6 +20,7 @@
 #include "commit-slab.h"
 #include "commit-graph.h"
 #include "commit-reach.h"
+#include "argv-array.h"
 
 static struct ref_msg {
 	const char *gone;
@@ -1790,21 +1791,62 @@ static int filter_pattern_match(struct ref_filter *filter, const char *refname)
 	return match_pattern(filter, refname);
 }
 
-/*
- * Find the longest prefix of pattern we can pass to
- * `for_each_fullref_in()`, namely the part of pattern preceding the
- * first glob character. (Note that `for_each_fullref_in()` is
- * perfectly happy working with a prefix that doesn't end at a
- * pathname component boundary.)
- */
-static void find_longest_prefix(struct strbuf *out, const char *pattern)
+static int qsort_strcmp(const void *va, const void *vb)
 {
-	const char *p;
+	const char *a = *(const char **)va;
+	const char *b = *(const char **)vb;
 
-	for (p = pattern; *p && !is_glob_special(*p); p++)
-		;
+	return strcmp(a, b);
+}
 
-	strbuf_add(out, pattern, p - pattern);
+static void find_longest_prefixes_1(struct string_list *out,
+				  struct strbuf *prefix,
+				  const char **patterns, size_t nr)
+{
+	size_t i;
+
+	for (i = 0; i < nr; i++) {
+		char c = patterns[i][prefix->len];
+		if (!c || is_glob_special(c)) {
+			string_list_append(out, prefix->buf);
+			return;
+		}
+	}
+
+	i = 0;
+	while (i < nr) {
+		size_t end;
+
+		/*
+		* Set "end" to the index of the element _after_ the last one
+		* in our group.
+		*/
+		for (end = i + 1; end < nr; end++) {
+			if (patterns[i][prefix->len] != patterns[end][prefix->len])
+				break;
+		}
+
+		strbuf_addch(prefix, patterns[i][prefix->len]);
+		find_longest_prefixes_1(out, prefix, patterns + i, end - i);
+		strbuf_setlen(prefix, prefix->len - 1);
+
+		i = end;
+	}
+}
+
+static void find_longest_prefixes(struct string_list *out,
+				  const char **patterns)
+{
+	struct argv_array sorted = ARGV_ARRAY_INIT;
+	struct strbuf prefix = STRBUF_INIT;
+
+	argv_array_pushv(&sorted, patterns);
+	QSORT(sorted.argv, sorted.argc, qsort_strcmp);
+
+	find_longest_prefixes_1(out, &prefix, sorted.argv, sorted.argc);
+
+	argv_array_clear(&sorted);
+	strbuf_release(&prefix);
 }
 
 /*
@@ -1817,7 +1859,8 @@ static int for_each_fullref_in_pattern(struct ref_filter *filter,
 				       void *cb_data,
 				       int broken)
 {
-	struct strbuf prefix = STRBUF_INIT;
+	struct string_list prefixes = STRING_LIST_INIT_DUP;
+	struct string_list_item *prefix;
 	int ret;
 
 	if (!filter->match_as_path) {
@@ -1843,21 +1886,15 @@ static int for_each_fullref_in_pattern(struct ref_filter *filter,
 		return for_each_fullref_in("", cb, cb_data, broken);
 	}
 
-	if (filter->name_patterns[1]) {
-		/*
-		 * multiple patterns; in theory this could still work as long
-		 * as the patterns are disjoint. We'd just make multiple calls
-		 * to for_each_ref(). But if they're not disjoint, we'd end up
-		 * reporting the same ref multiple times. So let's punt on that
-		 * for now.
-		 */
-		return for_each_fullref_in("", cb, cb_data, broken);
+	find_longest_prefixes(&prefixes, filter->name_patterns);
+
+	for_each_string_list_item(prefix, &prefixes) {
+		ret = for_each_fullref_in(prefix->string, cb, cb_data, broken);
+		if (ret)
+			break;
 	}
 
-	find_longest_prefix(&prefix, filter->name_patterns[0]);
-
-	ret = for_each_fullref_in(prefix.buf, cb, cb_data, broken);
-	strbuf_release(&prefix);
+	string_list_clear(&prefixes, 0);
 	return ret;
 }
 
diff --git a/t/t6300-for-each-ref.sh b/t/t6300-for-each-ref.sh
index d9235217fc..ab69aa176d 100755
--- a/t/t6300-for-each-ref.sh
+++ b/t/t6300-for-each-ref.sh
@@ -345,6 +345,32 @@ test_expect_success 'Verify descending sort' '
 	test_cmp expected actual
 '
 
+cat >expected <<\EOF
+refs/tags/testtag
+refs/tags/testtag-2
+EOF
+
+test_expect_success 'exercise patterns with prefixes' '
+	git tag testtag-2 &&
+	test_when_finished "git tag -d testtag-2" &&
+	git for-each-ref --format="%(refname)" \
+		refs/tags/testtag refs/tags/testtag-2 >actual &&
+	test_cmp expected actual
+'
+
+cat >expected <<\EOF
+refs/tags/testtag
+refs/tags/testtag-2
+EOF
+
+test_expect_success 'exercise glob patterns with prefixes' '
+	git tag testtag-2 &&
+	test_when_finished "git tag -d testtag-2" &&
+	git for-each-ref --format="%(refname)" \
+		refs/tags/testtag "refs/tags/testtag-*" >actual &&
+	test_cmp expected actual
+'
+
 cat >expected <<\EOF
 'refs/heads/master'
 'refs/remotes/origin/master'
-- 
2.21.0.203.g358da99528

Re: [PATCH 1/1] ref-filter.c: find disjoint pattern prefixes

[Jacob Keller]

On Wed, Jun 26, 2019 at 3:44 PM Taylor Blau <me@ttaylorr.com> wrote:
>
> Since cfe004a5a9 (ref-filter: limit traversal to prefix, 2017-05-22),
> the ref-filter code has sought to limit the traversals to a prefix of
> the given patterns.
>
> That code stopped short of handling more than one pattern, because it
> means invoking 'for_each_ref_in' multiple times. If we're not careful
> about which patterns overlap, we will output the same refs multiple
> times.

Right.

>
> For instance, consider the set of patterns 'refs/heads/a/*',
> 'refs/heads/a/b/c', and 'refs/tags/v1.0.0'. If we naďvely ran:
>
>   for_each_ref_in("refs/heads/a/*", ...);
>   for_each_ref_in("refs/heads/a/b/c", ...);
>   for_each_ref_in("refs/tags/v1.0.0", ...);
>
> we would see 'refs/heads/a/b/c' (and everything underneath it) twice.
>

Explaining why we ignored solving this before..

> Instead, we want to partition the patterns into disjoint sets, where we
> know that no ref will be matched by any two patterns in different sets.
> In the above, these are:
>
>   - {'refs/heads/a/*', 'refs/heads/a/b/c'}, and
>   - {'refs/tags/v1.0.0'}

Is this disjoint set calculation already existing, or did you have to
add it in this patch?

>
> Given one of these disjoint sets, what is a suitable pattern to pass to
> 'for_each_ref_in'? One approach is to compute the longest common prefix
> over all elements in that disjoint set, and let the caller cull out the
> refs they didn't want. Computing the longest prefix means that in most
> cases, we won't match too many things the caller would like to ignore.
>
> The longest common prefixes of the above are:
>
>   - {'refs/heads/a/*', 'refs/heads/a/b/c'} -> refs/heads/a/*
>   - {'refs/tags/v1.0.0'}                   -> refs/tags/v1.0.0
>
> We instead invoke:
>
>   for_each_ref_in("refs/heads/a/*", ...);
>   for_each_ref_in("refs/tags/v1.0.0", ...);
>
> Which provides us with the refs we were looking for with a minimal
> amount of extra cruft, but never a duplicate of the ref we asked for.
>
> Implemented here is an algorithm which accomplishes the above, which
> works as follows:
>
>   1. Lexicographically sort the given list of patterns.
>
>   2. Initialize 'prefix' to the empty string, where our goal is to
>      build each element in the above set of longest common prefixes.
>
>   3. Consider each pattern in the given set, and emit 'prefix' if it
>      reaches the end of a pattern, or touches a wildcard character. The
>      end of a string is treated as if it precedes a wildcard. (Note that
>      there is some room for future work to detect that, e.g., 'a?b' and
>      'abc' are disjoint).

Neat, so you're calculating disjoint patterns inline while also
figuring out which one is the "best" for any given pattern set...
Neat!

>
>   4. Otherwise, recurse on step (3) with the slice of the list
>      corresponding to our current prefix (i.e., the subset of patterns
>      that have our prefix as a literal string prefix.)
>
> This algorithm is 'O(kn + n log(n))', where 'k' is max(len(pattern)) for
> each pattern in the list, and 'n' is len(patterns).
>

ok, so if we can assume that k is some relatively small constant
number (since the maximum pattern length isn't likely to grow without
bounds), this is O(n*log(n)) on the number of patterns, so we don't
even approach n^2 even when we are given a large number of patterns.
Nice!

> By discovering this set of interesting patterns, we reduce the runtime
> of multi-pattern 'git for-each-ref' (and other ref traversals) from
> O(N) to O(n log(N)), where 'N' is the total number of packed references.

So here, n is the number of patterns still? This seems like a pretty
significant gane when we have a large number of packed references.

>
> Running 'git for-each-ref refs/tags/a refs/tags/b' on a repository with
> 10,000,000 refs in 'refs/tags/huge-N', my best-of-five times drop from:
>
>   real    0m5.805s
>   user    0m5.188s
>   sys     0m0.468s
>
> to:
>
>   real    0m0.001s
>   user    0m0.000s
>   sys     0m0.000s
>

That's a pretty significant decrease!

> On linux.git, the times to dig out two of the latest -rc tags drops from
> 0.002s to 0.001s, so the change on repositories with fewer tags is much
> less noticeable.
>

This explains why it might not have been done before.. many
repositories wouldn't benefit much.

That said, the patch description doesn't make it seem very
complicated. I did run out of time reading the message, so I'll have
to follow up reviewing the actual change below later. I think the
description of the goal and solution is sound though.

Thanks,
Jake

> Co-authored-by: Jeff King <peff@peff.net>
> Signed-off-by: Jeff King <peff@peff.net>
> Signed-off-by: Taylor Blau <me@ttaylorr.com>
> ---
>  ref-filter.c            | 89 +++++++++++++++++++++++++++++------------
>  t/t6300-for-each-ref.sh | 26 ++++++++++++
>  2 files changed, 89 insertions(+), 26 deletions(-)
>
> diff --git a/ref-filter.c b/ref-filter.c
> index 8500671bc6..3e10fd647b 100644
> --- a/ref-filter.c
> +++ b/ref-filter.c
> @@ -20,6 +20,7 @@
>  #include "commit-slab.h"
>  #include "commit-graph.h"
>  #include "commit-reach.h"
> +#include "argv-array.h"
>
>  static struct ref_msg {
>         const char *gone;
> @@ -1790,21 +1791,62 @@ static int filter_pattern_match(struct ref_filter *filter, const char *refname)
>         return match_pattern(filter, refname);
>  }
>
> -/*
> - * Find the longest prefix of pattern we can pass to
> - * `for_each_fullref_in()`, namely the part of pattern preceding the
> - * first glob character. (Note that `for_each_fullref_in()` is
> - * perfectly happy working with a prefix that doesn't end at a
> - * pathname component boundary.)
> - */
> -static void find_longest_prefix(struct strbuf *out, const char *pattern)
> +static int qsort_strcmp(const void *va, const void *vb)
>  {
> -       const char *p;
> +       const char *a = *(const char **)va;
> +       const char *b = *(const char **)vb;
>
> -       for (p = pattern; *p && !is_glob_special(*p); p++)
> -               ;
> +       return strcmp(a, b);
> +}
>
> -       strbuf_add(out, pattern, p - pattern);
> +static void find_longest_prefixes_1(struct string_list *out,
> +                                 struct strbuf *prefix,
> +                                 const char **patterns, size_t nr)
> +{
> +       size_t i;
> +
> +       for (i = 0; i < nr; i++) {
> +               char c = patterns[i][prefix->len];
> +               if (!c || is_glob_special(c)) {
> +                       string_list_append(out, prefix->buf);
> +                       return;
> +               }
> +       }
> +
> +       i = 0;
> +       while (i < nr) {
> +               size_t end;
> +
> +               /*
> +               * Set "end" to the index of the element _after_ the last one
> +               * in our group.
> +               */
> +               for (end = i + 1; end < nr; end++) {
> +                       if (patterns[i][prefix->len] != patterns[end][prefix->len])
> +                               break;
> +               }
> +
> +               strbuf_addch(prefix, patterns[i][prefix->len]);
> +               find_longest_prefixes_1(out, prefix, patterns + i, end - i);
> +               strbuf_setlen(prefix, prefix->len - 1);
> +
> +               i = end;
> +       }
> +}
> +
> +static void find_longest_prefixes(struct string_list *out,
> +                                 const char **patterns)
> +{
> +       struct argv_array sorted = ARGV_ARRAY_INIT;
> +       struct strbuf prefix = STRBUF_INIT;
> +
> +       argv_array_pushv(&sorted, patterns);
> +       QSORT(sorted.argv, sorted.argc, qsort_strcmp);
> +
> +       find_longest_prefixes_1(out, &prefix, sorted.argv, sorted.argc);
> +
> +       argv_array_clear(&sorted);
> +       strbuf_release(&prefix);
>  }
>
>  /*
> @@ -1817,7 +1859,8 @@ static int for_each_fullref_in_pattern(struct ref_filter *filter,
>                                        void *cb_data,
>                                        int broken)
>  {
> -       struct strbuf prefix = STRBUF_INIT;
> +       struct string_list prefixes = STRING_LIST_INIT_DUP;
> +       struct string_list_item *prefix;
>         int ret;
>
>         if (!filter->match_as_path) {
> @@ -1843,21 +1886,15 @@ static int for_each_fullref_in_pattern(struct ref_filter *filter,
>                 return for_each_fullref_in("", cb, cb_data, broken);
>         }
>
> -       if (filter->name_patterns[1]) {
> -               /*
> -                * multiple patterns; in theory this could still work as long
> -                * as the patterns are disjoint. We'd just make multiple calls
> -                * to for_each_ref(). But if they're not disjoint, we'd end up
> -                * reporting the same ref multiple times. So let's punt on that
> -                * for now.
> -                */
> -               return for_each_fullref_in("", cb, cb_data, broken);
> +       find_longest_prefixes(&prefixes, filter->name_patterns);
> +
> +       for_each_string_list_item(prefix, &prefixes) {
> +               ret = for_each_fullref_in(prefix->string, cb, cb_data, broken);
> +               if (ret)
> +                       break;
>         }
>
> -       find_longest_prefix(&prefix, filter->name_patterns[0]);
> -
> -       ret = for_each_fullref_in(prefix.buf, cb, cb_data, broken);
> -       strbuf_release(&prefix);
> +       string_list_clear(&prefixes, 0);
>         return ret;
>  }
>
> diff --git a/t/t6300-for-each-ref.sh b/t/t6300-for-each-ref.sh
> index d9235217fc..ab69aa176d 100755
> --- a/t/t6300-for-each-ref.sh
> +++ b/t/t6300-for-each-ref.sh
> @@ -345,6 +345,32 @@ test_expect_success 'Verify descending sort' '
>         test_cmp expected actual
>  '
>
> +cat >expected <<\EOF
> +refs/tags/testtag
> +refs/tags/testtag-2
> +EOF
> +
> +test_expect_success 'exercise patterns with prefixes' '
> +       git tag testtag-2 &&
> +       test_when_finished "git tag -d testtag-2" &&
> +       git for-each-ref --format="%(refname)" \
> +               refs/tags/testtag refs/tags/testtag-2 >actual &&
> +       test_cmp expected actual
> +'
> +
> +cat >expected <<\EOF
> +refs/tags/testtag
> +refs/tags/testtag-2
> +EOF
> +
> +test_expect_success 'exercise glob patterns with prefixes' '
> +       git tag testtag-2 &&
> +       test_when_finished "git tag -d testtag-2" &&
> +       git for-each-ref --format="%(refname)" \
> +               refs/tags/testtag "refs/tags/testtag-*" >actual &&
> +       test_cmp expected actual
> +'
> +
>  cat >expected <<\EOF
>  'refs/heads/master'
>  'refs/remotes/origin/master'
> --
> 2.21.0.203.g358da99528

Re: [PATCH 1/1] ref-filter.c: find disjoint pattern prefixes

[Jacob Keller]

On Wed, Jun 26, 2019 at 3:44 PM Taylor Blau <me@ttaylorr.com> wrote:
>

Ok, now I've got some time to look at the code itself.

>  ref-filter.c            | 89 +++++++++++++++++++++++++++++------------
>  t/t6300-for-each-ref.sh | 26 ++++++++++++
>  2 files changed, 89 insertions(+), 26 deletions(-)
>
> diff --git a/ref-filter.c b/ref-filter.c
> index 8500671bc6..3e10fd647b 100644
> --- a/ref-filter.c
> +++ b/ref-filter.c
> @@ -20,6 +20,7 @@
>  #include "commit-slab.h"
>  #include "commit-graph.h"
>  #include "commit-reach.h"
> +#include "argv-array.h"
>
>  static struct ref_msg {
>         const char *gone;
> @@ -1790,21 +1791,62 @@ static int filter_pattern_match(struct ref_filter *filter, const char *refname)
>         return match_pattern(filter, refname);
>  }
>
> -/*
> - * Find the longest prefix of pattern we can pass to
> - * `for_each_fullref_in()`, namely the part of pattern preceding the
> - * first glob character. (Note that `for_each_fullref_in()` is
> - * perfectly happy working with a prefix that doesn't end at a
> - * pathname component boundary.)
> - */
> -static void find_longest_prefix(struct strbuf *out, const char *pattern)
> +static int qsort_strcmp(const void *va, const void *vb)
>  {
> -       const char *p;
> +       const char *a = *(const char **)va;
> +       const char *b = *(const char **)vb;
>
> -       for (p = pattern; *p && !is_glob_special(*p); p++)
> -               ;
> +       return strcmp(a, b);
> +}
>
> -       strbuf_add(out, pattern, p - pattern);
> +static void find_longest_prefixes_1(struct string_list *out,
> +                                 struct strbuf *prefix,
> +                                 const char **patterns, size_t nr)
> +{
> +       size_t i;
> +
> +       for (i = 0; i < nr; i++) {
> +               char c = patterns[i][prefix->len];
> +               if (!c || is_glob_special(c)) {
> +                       string_list_append(out, prefix->buf);
> +                       return;
> +               }
> +       }

Ok, so we loop over the patterns, find the last character up to our
current prefix length, and check if it's either the end of the string,
or a special glob character. I assume that the patterns are sorted so
that prefix->len never goes past the array?

If we find one, we just add this to the list and return, since we
found an unambigous one.

Otherwise, we keep searching recursively.

So, prefix is a strbuf, and its length will initially be zero. So, we
check all patterns, up to the prefix length and check the character
just past the end of our prefix. If it matches a NUL or glob
character, then we have found an exact match up to a glob, so this
gets returned.

Otherwise we continue:

> +
> +       i = 0;
> +       while (i < nr) {
> +               size_t end;
> +

Here, we're going to loop from beginning to end of all of the strings.

> +               /*
> +               * Set "end" to the index of the element _after_ the last one
> +               * in our group.
> +               */
> +               for (end = i + 1; end < nr; end++) {
> +                       if (patterns[i][prefix->len] != patterns[end][prefix->len])
> +                               break;
> +               }
> +

We break on the first string which doesn't have the same length as our
current prefix, but start with the ones after the current loop
iteration.

> +               strbuf_addch(prefix, patterns[i][prefix->len]);
> +               find_longest_prefixes_1(out, prefix, patterns + i, end - i);
> +               strbuf_setlen(prefix, prefix->len - 1);
> +

We'll add the next character to the prefix, and then find longest
prefixes again.

This basically has us recurse and keep adding additional characters,
essentially splitting the strings apart by their disjoint sets.

I think this works, but it's definitely not clear from reading exactly
what is going on. I think this algorithm would benefit from a comment,
since it doesn't quite seem to match your description in the commit
message.

> +               i = end;
> +       }
> +}
> +
> +static void find_longest_prefixes(struct string_list *out,
> +                                 const char **patterns)
> +{
> +       struct argv_array sorted = ARGV_ARRAY_INIT;
> +       struct strbuf prefix = STRBUF_INIT;
> +
> +       argv_array_pushv(&sorted, patterns);
> +       QSORT(sorted.argv, sorted.argc, qsort_strcmp);
> +

We've sorted the patterns, and then we call find_longest_prefixes_1. Ok.

> +       find_longest_prefixes_1(out, &prefix, sorted.argv, sorted.argc);
> +
> +       argv_array_clear(&sorted);
> +       strbuf_release(&prefix);
>  }
>
>  /*
> @@ -1817,7 +1859,8 @@ static int for_each_fullref_in_pattern(struct ref_filter *filter,
>                                        void *cb_data,
>                                        int broken)
>  {
> -       struct strbuf prefix = STRBUF_INIT;
> +       struct string_list prefixes = STRING_LIST_INIT_DUP;
> +       struct string_list_item *prefix;
>         int ret;
>
>         if (!filter->match_as_path) {
> @@ -1843,21 +1886,15 @@ static int for_each_fullref_in_pattern(struct ref_filter *filter,
>                 return for_each_fullref_in("", cb, cb_data, broken);
>         }
>
> -       if (filter->name_patterns[1]) {
> -               /*
> -                * multiple patterns; in theory this could still work as long
> -                * as the patterns are disjoint. We'd just make multiple calls
> -                * to for_each_ref(). But if they're not disjoint, we'd end up
> -                * reporting the same ref multiple times. So let's punt on that
> -                * for now.
> -                */
> -               return for_each_fullref_in("", cb, cb_data, broken);
> +       find_longest_prefixes(&prefixes, filter->name_patterns);
> +
> +       for_each_string_list_item(prefix, &prefixes) {
> +               ret = for_each_fullref_in(prefix->string, cb, cb_data, broken);
> +               if (ret)
> +                       break;
>         }
>
> -       find_longest_prefix(&prefix, filter->name_patterns[0]);
> -
> -       ret = for_each_fullref_in(prefix.buf, cb, cb_data, broken);
> -       strbuf_release(&prefix);
> +       string_list_clear(&prefixes, 0);
>         return ret;
>  }
>
> diff --git a/t/t6300-for-each-ref.sh b/t/t6300-for-each-ref.sh
> index d9235217fc..ab69aa176d 100755
> --- a/t/t6300-for-each-ref.sh
> +++ b/t/t6300-for-each-ref.sh
> @@ -345,6 +345,32 @@ test_expect_success 'Verify descending sort' '
>         test_cmp expected actual
>  '
>
> +cat >expected <<\EOF
> +refs/tags/testtag
> +refs/tags/testtag-2
> +EOF
> +
> +test_expect_success 'exercise patterns with prefixes' '
> +       git tag testtag-2 &&
> +       test_when_finished "git tag -d testtag-2" &&
> +       git for-each-ref --format="%(refname)" \
> +               refs/tags/testtag refs/tags/testtag-2 >actual &&
> +       test_cmp expected actual
> +'
> +
> +cat >expected <<\EOF
> +refs/tags/testtag
> +refs/tags/testtag-2
> +EOF
> +
> +test_expect_success 'exercise glob patterns with prefixes' '
> +       git tag testtag-2 &&
> +       test_when_finished "git tag -d testtag-2" &&
> +       git for-each-ref --format="%(refname)" \
> +               refs/tags/testtag "refs/tags/testtag-*" >actual &&
> +       test_cmp expected actual
> +'
> +
>  cat >expected <<\EOF
>  'refs/heads/master'
>  'refs/remotes/origin/master'
> --
> 2.21.0.203.g358da99528

Re: [PATCH 1/1] ref-filter.c: find disjoint pattern prefixes

[Taylor Blau]

Hi Jacob,

On Wed, Jun 26, 2019 at 05:37:42PM -0700, Jacob Keller wrote:
> [ ... ]
>
> > Instead, we want to partition the patterns into disjoint sets, where we
> > know that no ref will be matched by any two patterns in different sets.
> > In the above, these are:
> >
> >   - {'refs/heads/a/*', 'refs/heads/a/b/c'}, and
> >   - {'refs/tags/v1.0.0'}
>
> Is this disjoint set calculation already existing, or did you have to
> add it in this patch?

Both the disjoint set calculation and the prefixing procedure are new in
this patch. But, we're never actually computing this disjoint set
explicitly, rather, we build it up implicitly while computing what will
become the longest prefixes of each subset.

> >   4. Otherwise, recurse on step (3) with the slice of the list
> >      corresponding to our current prefix (i.e., the subset of patterns
> >      that have our prefix as a literal string prefix.)
> >
> > This algorithm is 'O(kn + n log(n))', where 'k' is max(len(pattern)) for
> > each pattern in the list, and 'n' is len(patterns).
> >
>
> ok, so if we can assume that k is some relatively small constant
> number (since the maximum pattern length isn't likely to grow without
> bounds), this is O(n*log(n)) on the number of patterns, so we don't
> even approach n^2 even when we are given a large number of patterns.
> Nice!
>
> > By discovering this set of interesting patterns, we reduce the runtime
> > of multi-pattern 'git for-each-ref' (and other ref traversals) from
> > O(N) to O(n log(N)), where 'N' is the total number of packed references.
>
> So here, n is the number of patterns still? This seems like a pretty
> significant gane when we have a large number of packed references.

Yes, 'n' is the number of patterns given. For e.g., the invocation

  $ git for-each-ref 'refs/heads/*' 'refs/tags/*'

has 'n = 2', and 'N' is unknown. The asymptotics here are really
comparing the case where we previously didn't make any effort to compute
good queries, and resorted to a linear scan of all packed references,
compared to now where we have at most one query per pattern, resulting
in a logarithmic-time scan of .git/packed-refs.

> >
> > Running 'git for-each-ref refs/tags/a refs/tags/b' on a repository with
> > 10,000,000 refs in 'refs/tags/huge-N', my best-of-five times drop from:
> >
> >   real    0m5.805s
> >   user    0m5.188s
> >   sys     0m0.468s
> >
> > to:
> >
> >   real    0m0.001s
> >   user    0m0.000s
> >   sys     0m0.000s
> >
>
> That's a pretty significant decrease!

Yes, it's quite good here, but it's designed to be that way ;-). Like I
note below, the real world speed-ups aren't quite as remarkable, but
it's not uncommon for us at GitHub to have a repository of the above
shape in terms of the number of references.

So, it's an increase almost no matter where you are, but it works
especially well for us.

> > On linux.git, the times to dig out two of the latest -rc tags drops from
> > 0.002s to 0.001s, so the change on repositories with fewer tags is much
> > less noticeable.
> >
>
> This explains why it might not have been done before.. many
> repositories wouldn't benefit much.
>
> That said, the patch description doesn't make it seem very
> complicated. I did run out of time reading the message, so I'll have
> to follow up reviewing the actual change below later. I think the
> description of the goal and solution is sound though.

Thanks for the initial review :-).

Thanks,
Taylor

Re: [PATCH 1/1] ref-filter.c: find disjoint pattern prefixes

[Taylor Blau]

Hi Jacob,

On Thu, Jun 27, 2019 at 04:21:57PM -0700, Jacob Keller wrote:
> On Wed, Jun 26, 2019 at 3:44 PM Taylor Blau <me@ttaylorr.com> wrote:
> >
>
> Ok, now I've got some time to look at the code itself.
>
> >  ref-filter.c            | 89 +++++++++++++++++++++++++++++------------
> >  t/t6300-for-each-ref.sh | 26 ++++++++++++
> >  2 files changed, 89 insertions(+), 26 deletions(-)
> >
> > diff --git a/ref-filter.c b/ref-filter.c
> > index 8500671bc6..3e10fd647b 100644
> > --- a/ref-filter.c
> > +++ b/ref-filter.c
> > @@ -20,6 +20,7 @@
> >  #include "commit-slab.h"
> >  #include "commit-graph.h"
> >  #include "commit-reach.h"
> > +#include "argv-array.h"
> >
> >  static struct ref_msg {
> >         const char *gone;
> > @@ -1790,21 +1791,62 @@ static int filter_pattern_match(struct ref_filter *filter, const char *refname)
> >         return match_pattern(filter, refname);
> >  }
> >
> > -/*
> > - * Find the longest prefix of pattern we can pass to
> > - * `for_each_fullref_in()`, namely the part of pattern preceding the
> > - * first glob character. (Note that `for_each_fullref_in()` is
> > - * perfectly happy working with a prefix that doesn't end at a
> > - * pathname component boundary.)
> > - */
> > -static void find_longest_prefix(struct strbuf *out, const char *pattern)
> > +static int qsort_strcmp(const void *va, const void *vb)
> >  {
> > -       const char *p;
> > +       const char *a = *(const char **)va;
> > +       const char *b = *(const char **)vb;
> >
> > -       for (p = pattern; *p && !is_glob_special(*p); p++)
> > -               ;
> > +       return strcmp(a, b);
> > +}
> >
> > -       strbuf_add(out, pattern, p - pattern);
> > +static void find_longest_prefixes_1(struct string_list *out,
> > +                                 struct strbuf *prefix,
> > +                                 const char **patterns, size_t nr)
> > +{
> > +       size_t i;
> > +
> > +       for (i = 0; i < nr; i++) {
> > +               char c = patterns[i][prefix->len];
> > +               if (!c || is_glob_special(c)) {
> > +                       string_list_append(out, prefix->buf);
> > +                       return;
> > +               }
> > +       }
>
> Ok, so we loop over the patterns, find the last character up to our
> current prefix length, and check if it's either the end of the string,
> or a special glob character. I assume that the patterns are sorted so
> that prefix->len never goes past the array?

That's right. The idea here is that if we compare two patterns character
by character, we can detect where the references "split", and thus
constitute membership of unique disjoint sets.

Consider the following example:

  refs/heads/a
  refs/heads/b
             ^

Here, we were given, say 'git for-each-ref refs/heads/a refs/heads/b',
and appropriate queries are:

  - '*'
  - 'refs/*'
  - 'refs/heads/*'
  - 'refs/heads/a' with 'refs/heads/b'

When we have advanced our cursor up until the position where the
patterns read 'a', and 'b', respectively, we have built up so far as our
pattern: 'refs/heads/'. We would like to realize that those two can
create two queries, so when we discover that 'a' != 'b', we "split" the
strbuf into two patterns in the next level of recursion.

One alternative view is that we are trying to construct the shortest
paths from the root of the tree to the leaf 'NUL' nodes as below:

                   a --- NUL
                 /
  refs --- heads
                 \
                   b --- NUL

> If we find one, we just add this to the list and return, since we
> found an unambigous one.
>
> Otherwise, we keep searching recursively.
>
> So, prefix is a strbuf, and its length will initially be zero. So, we
> check all patterns, up to the prefix length and check the character
> just past the end of our prefix. If it matches a NUL or glob
> character, then we have found an exact match up to a glob, so this
> gets returned.

Right: two additional wrinkles here.

One is that we don't have very careful handling of wildcard characters,
i.e., between 'a*c' and '*bc' we give up and split those at 'a*', and
'*b' (an equally good query would be '**c'), so treat the presence of a
wildcard character the same way we would a disagreement in character as
above.

The second wrinkle is that we originally wrote this patch looking for
disagreements in _ref component_, i.e., that given 'refs/heads/a' with
'refs/tags/a', we would split because 'heads != tags', not 'h != t'. The
references backend (to my surprise) does per-character matching, so
that's why we do it, too.

> Here, we're going to loop from beginning to end of all of the strings.
>
> > +               /*
> > +               * Set "end" to the index of the element _after_ the last one
> > +               * in our group.
> > +               */
> > +               for (end = i + 1; end < nr; end++) {
> > +                       if (patterns[i][prefix->len] != patterns[end][prefix->len])
> > +                               break;
> > +               }
> > +
>
> We break on the first string which doesn't have the same length as our
> current prefix, but start with the ones after the current loop
> iteration.
>
> > +               strbuf_addch(prefix, patterns[i][prefix->len]);
> > +               find_longest_prefixes_1(out, prefix, patterns + i, end - i);
> > +               strbuf_setlen(prefix, prefix->len - 1);
> > +

Right, we have a potentially long list of patterns underneath that we
have just covered, so we can skip those.

> We'll add the next character to the prefix, and then find longest
> prefixes again.
>
> This basically has us recurse and keep adding additional characters,
> essentially splitting the strings apart by their disjoint sets.
>
> I think this works, but it's definitely not clear from reading exactly
> what is going on. I think this algorithm would benefit from a comment,
> since it doesn't quite seem to match your description in the commit
> message.
> > +       find_longest_prefixes_1(out, &prefix, sorted.argv, sorted.argc);
> > +
> > +       argv_array_clear(&sorted);
> > +       strbuf_release(&prefix);

Fair enough, I'm happy to add a comment if others like, too.

Thanks,
Taylor