fsmonitor: fix watchman integration

When running Git commands quickly -- such as in a shell script or the test suite -- the Git commands frequently complete and start again during the same second. The example fsmonitor hooks to integrate with Watchman truncate the nanosecond times to seconds. In principle, this is fine, as Watchman claims to use inclusive comparisons [1]. The result should only be an over-representation of the changed paths since the last Git command.

However, Watchman's own documentation claims "Using a timestamp is prone to race conditions in understanding the complete state of the file tree" [2]. All of their documented examples use a "clockspec" that looks like 'c:123:234'. Git should eventually learn how to store this type of string to provide a stronger integration, but that will be a more invasive change.

When using GIT_TEST_FSMONITOR="$(pwd)/t7519/fsmonitor-watchman", scripts such as t7519-wtstatus.sh fail due to these race conditions. In fact, running any test script with GIT_TEST_FSMONITOR pointing at t/t7519/fsmonitor-wathcman will cause failures in the test_commit function. The 'git add "$indir$file"' command fails due to not enough time between the creation of '$file' and the 'git add' command.

For now, subtract one second from the timestamp we pass to Watchman. This will make our window large enough to avoid these race conditions. Increasing the window causes tests like t7519-wtstatus.sh to pass.

When the integration was introduced in def437671 (fsmonitor: add a sample integration script for Watchman, 2018-09-22), the query included an expression that would ignore files created and deleted in that window. The performance reason for this change was to ignore temporary files created by a build between Git commands. However, this causes failures in script scenarios where Git is creating or deleting files quickly.

When using GIT_TEST_FSMONITOR as before, t2203-add-intent.sh fails due to this add-and-delete race condition.

By removing the "expression" from the Watchman query, we remove this race condition. It will lead to some performance degradation in the case of users creating and deleting temporary files inside their working directory between Git commands. However, that is a cost we need to pay to be correct.

[1] https://github.com/facebook/watchman/blob/master/query/since.cpp#L35-L39
[2] https://facebook.github.io/watchman/docs/clockspec.html

Kevin Willford (1):
fsmonitor: fix watchman integration

 t/t7519/fsmonitor-watchman                 | 13 ++++---------
 templates/hooks--fsmonitor-watchman.sample | 13 ++++---------
 2 files changed, 8 insertions(+), 18 deletions(-)

base-commit: da72936f544fec5a335e66432610e4cef4430991
Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-444%2Fkewillford%2Ffsmonitor-watchman-fix-v1
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-444/kewillford/fsmonitor-watchman-fix-v1
Pull-Request: https://github.com/gitgitgadget/git/pull/444

gitgitgadget

From: Kevin Willford <[email protected]>

When running Git commands quickly -- such as in a shell script or the test suite -- the Git commands frequently complete and start again during the same second. The example fsmonitor hooks to integrate with Watchman truncate the nanosecond times to seconds. In principle, this is fine, as Watchman claims to use inclusive comparisons [1]. The result should only be an over-representation of the changed paths since the last Git command.

However, Watchman's own documentation claims "Using a timestamp is prone to race conditions in understanding the complete state of the file tree" [2]. All of their documented examples use a "clockspec" that looks like 'c:123:234'. Git should eventually learn how to store this type of string to provide a stronger integration, but that will be a more invasive change.

When using GIT_TEST_FSMONITOR="$(pwd)/t7519/fsmonitor-watchman", scripts such as t7519-wtstatus.sh fail due to these race conditions. In fact, running any test script with GIT_TEST_FSMONITOR pointing at t/t7519/fsmonitor-wathcman will cause failures in the test_commit function. The 'git add "$indir$file"' command fails due to not enough time between the creation of '$file' and the 'git add' command.

For now, subtract one second from the timestamp we pass to Watchman. This will make our window large enough to avoid these race conditions. Increasing the window causes tests like t7519-wtstatus.sh to pass.

When the integration was introduced in def437671 (fsmonitor: add a sample integration script for Watchman, 2018-09-22), the query included an expression that would ignore files created and deleted in that window. The performance reason for this change was to ignore temporary files created by a build between Git commands. However, this causes failures in script scenarios where Git is creating or deleting files quickly.

When using GIT_TEST_FSMONITOR as before, t2203-add-intent.sh fails due to this add-and-delete race condition.

By removing the "expression" from the Watchman query, we remove this race condition. It will lead to some performance degradation in the case of users creating and deleting temporary files inside their working directory between Git commands. However, that is a cost we need to pay to be correct.

[1] https://github.com/facebook/watchman/blob/master/query/since.cpp#L35-L39
[2] https://facebook.github.io/watchman/docs/clockspec.html

---
 t/t7519/fsmonitor-watchman                 | 13 ++++---------
 templates/hooks--fsmonitor-watchman.sample | 13 ++++---------
 2 files changed, 8 insertions(+), 18 deletions(-)

index 5514edcf68..d8e7a1e5ba 100755
@@ -23,7 +23,8 @@ my ($version, $time) = @ARGV;
 
 if ($version == 1) {
 	# convert nanoseconds to seconds
-	$time = int $time / 1000000000;
+	# subtract one second to make sure watchman will return all changes
+	$time = int ($time / 1000000000) - 1;
 } else {
 	die "Unsupported query-fsmonitor hook version '$version'.\n" .
 	    "Falling back to scanning...\n";

@@ -54,18 +55,12 @@ sub launch_watchman {
 	#
 	# To accomplish this, we're using the "since" generator to use the
 	# recency index to select candidate nodes and "fields" to limit the
-	# output to file names only. Then we're using the "expression" term to
-	# further constrain the results.
-	#
-	# The category of transient files that we want to ignore will have a
-	# creation clock (cclock) newer than $time_t value and will also not
-	# currently exist.
+	# output to file names only.
 
 	my $query = <<"	END";
 		["query", "$git_work_tree", {
 			"since": $time,
-			"fields": ["name"],
-			"expression": ["not", ["allof", ["since", $time, "cclock"], ["not", "exists"]]]
+			"fields": ["name"]
 		}]
 	END
 	

index e673bb3980..ef94fa2938 100755
@@ -22,7 +22,8 @@ my ($version, $time) = @ARGV;
 
 if ($version == 1) {
 	# convert nanoseconds to seconds
-	$time = int $time / 1000000000;
+	# subtract one second to make sure watchman will return all changes
+	$time = int ($time / 1000000000) - 1;
 } else {
 	die "Unsupported query-fsmonitor hook version '$version'.\n" .
 	    "Falling back to scanning...\n";

@@ -53,18 +54,12 @@ sub launch_watchman {
 	#
 	# To accomplish this, we're using the "since" generator to use the
 	# recency index to select candidate nodes and "fields" to limit the
-	# output to file names only. Then we're using the "expression" term to
-	# further constrain the results.
-	#
-	# The category of transient files that we want to ignore will have a
-	# creation clock (cclock) newer than $time_t value and will also not
-	# currently exist.
+	# output to file names only.
 
 	my $query = <<"	END";
 		["query", "$git_work_tree", {
 			"since": $time,
-			"fields": ["name"],
-			"expression": ["not", ["allof", ["since", $time, "cclock"], ["not", "exists"]]]
+			"fields": ["name"]
 		}]
 	END
 

gitgitgadget

So, it doesn't seem to use "inclusive" and we need a workaround?

Thanks, will queue.

That is what is seems like. I would like to dig into the watchman code to understand what is really going on. They also document that "Using a timestamp is prone to race conditions in understanding the complete state of the file tree." Which could be the cause since the tests are running things in quick succession, i.e. change a file, run a git command.

Long term we should switch to using watchman's clock id which the documentation says does not have the race conditions. But the clock id is a string and would take more invasive changes to integrate that into the index where we are now simply using a uint64_t.

I should mention that I'm working on a patch series that will allow us to use GIT_TEST_FSMONITOR pointing at Watchman in our CI builds. There are several places where our fsmonitor integration does not work well (such as when we delete submodules), but most of the remaining fixes are small.

This timing issue fixes MOST of the problems we see when running the test suite.

Thanks,
-Stolee

Nit: please refer to the specific blob in this link. The file's content in 'master' might change in time, and then the highlight will point to different lines. Worse, the file might be renamed, and then we'll get a broken link.