Summary:
Ref T13552. Currently, commit lists are sometimes rendered as an object list and sometimes rendered as a table. There are two separate views for table rendering.
Add a fourth view ("list, with a graph") with the eventual intent of unifying all the other views. For now, this only replaces "HistoryListView" -- and needs some more work to really be a convincing replacement.
Test Plan:
- Looked at "History" in Diffusion, saw an ugly view with all the information we want.
- Grepped for "HistoryListView", no hits.
Maniphest Tasks: T13552
Differential Revision: https://secure.phabricator.com/D21410
Summary:
Ref T13552. Currently, Diffusion has two effectively identical history views, the "Graph" view and the "History" view.
These arose out of product uncertainty about the importance of the graph, but I think we can just put the graph on the "object item list" view and merge these views.
Test Plan: Looked at repositories in Diffusion, no longer saw a "Graph" tab. Grepped for "graph"-related symbols.
Maniphest Tasks: T13552
Differential Revision: https://secure.phabricator.com/D21409
Summary:
Ref T13552. Currently, the repository landing page has a panel with recent commits. This is accessible by clicking "History" and usually below the fold, so it's not clearly useful.
Since I'm consolidating this code anyway to fix an issue with the import pipeline, just get rid of this history view.
Test Plan: Viewed a repository landing page, no longer saw a history panel.
Maniphest Tasks: T13552
Differential Revision: https://secure.phabricator.com/D21408
Summary: Ref T13552. This older class has no callers; tag and branch listings were replaced with an "ObjectList" view.
Test Plan: Grepped for "DiffusionTagTableView", got no hits.
Maniphest Tasks: T13552
Differential Revision: https://secure.phabricator.com/D21407
Summary:
Ref T13552. I'm trying to reduce the number of direct callers to commit authorship metadata. This header seems low-value enough to simply remove; this information is shown more clearly and prominently in the "Provenance" UI.
In particular, commits have multiple dates (authored, committed, pushed) but this header shows only one. It currently shows the author identity and the commit date, which isn't entirely correct. And it potentially uses an "Identity" as a timeline actor, which is conceptually fine but not entirely firm ground.
Test Plan: Viewed a commit, saw no more subheader.
Maniphest Tasks: T13552
Differential Revision: https://secure.phabricator.com/D21406
Summary:
Ref T13552. Give "Commit" objects a more modern, identity-aware way to render author and committer information.
This uses handles in a more modern way and gives us a single read callsite for raw author and committer names.
Test Plan:
- Grepped for callers to the old methods, found none. (There are a lot of "renderAuthor()" callers in transactions, but this call takes no parameters.)
- Viewed some commits, saw sensible lists of authors and committers.
Maniphest Tasks: T13552
Differential Revision: https://secure.phabricator.com/D21405
Summary:
Ref T13552. When viewing a directory in Diffusion, we make an Ajax call to get the last commit for each path.
This call currently pulls author information, since an older version of this UI showed author information.
The current UI does not show author information, so this parameter is unused. Delete the code which builds it.
Test Plan: Grepped for `'author'` and references to the "pull-lastmodified" behavior. This behavior is invoked in only one place, which never generates an author placeholder.
Maniphest Tasks: T13552
Differential Revision: https://secure.phabricator.com/D21404
Summary:
See T13564. In Chrome only, printing tables with a cell containing an unbroken monospaced text element fails to wrap/break the cell.
Adding "overflow-wrap" appears to fix this without making anything worse. Try this until new problems arise.
Test Plan: Printed such a table to PDF in Chrome, got wrapping with all content visible in the PDF.
Differential Revision: https://secure.phabricator.com/D21439
Summary:
See PHI1834. It's not obvious why this "+1" is present in the code, but it causes inlines to be adjusted incorrectly when a file is not modified across changes. See D21435.
Remove it, which appears to produce accurate adjustment behavior.
Test Plan:
- See D21435 for instructions to build a change, where a file with lines "A-Z" is unmodified across Diff 1 and Diff 2.
- Left inlines on lines 14, 17-19, and 16-26 (end of the file) on Diff 1.
- Before: saw inlines incorrectly adjusted to lines 15, 18, and 17 on Diff 2. Before D21435, the last inline was culled by the rendering engine.
- After: saw inlines correctly adjusted to lines 14, 17, and 16 (the same lines as the original), render properly, and highlight the correct lines when hovered.
Differential Revision: https://secure.phabricator.com/D21436
Summary:
See PHI1834. Currently, the inline adjustment engine can sometime "adjust" an inline off the end of a diff. If it does, we lay it out on an invalid display line here and never render it.
Instead, make sure that layout never puts a comment on an invalid line, so the UI is robust against questionable decisions by the adjustment engine: no adjustment should be able to accidentally discard an inline.
Test Plan:
- Created a two diff revision, where Diffs 1 and 2 have "alphabet.txt" with A-Z on one line each. The file is unchanged across diffs; some other file is changed.
- Added a comment to lines P-Z of Diff 1.
- Before: comment is adjusted out of range on Diff 2 and not shown in the UI.
- After: comment is still adjusted out of range internally, but now corrected into the display range and shown.
Differential Revision: https://secure.phabricator.com/D21435
Summary:
Ref PHI1835. Generally, Jupyter notebooks in the wild may store source and markdown content as either a single string or a list of strings.
Make the renderer read these formats more consistently. In particular, this fixes rendering of code blocks stored as a single string.
This also fixes an issue where cell labels were double-rendered in diff views.
Test Plan:
Created a notebook with a code block represented on disk as a single string, rendered a diff from it.
{F7696071}
Differential Revision: https://secure.phabricator.com/D21434
Summary:
See PHI1839. Currently, the "No newline at end of file" text is dropped in the 1-up diff view for changes that affect a file with no trailing newline.
Track it through the construction of diff primitivies more carefully.
Test Plan: {F7695760}
Differential Revision: https://secure.phabricator.com/D21433
Summary:
Although I'm not entirely thrilled about doing flow control like this (as an actual action in a build plan), I believe this build step works correctly and there's no fancy replacement mechanism on the immediate horizon, and this didn't send us down a slippery slope of Turing-complete builds encoded without real structure or context. Just kick it out of prototype.
(Other approaches which might be better in the long run are things like "this is a top-level behavior on the build plan itself" and/or "build plans are written in a DSL, not a Javascript UI".)
Test Plan: Added a new build step, saw this as an option in the "Flow Control" section.
Subscribers: PHID-OPKG-gm6ozazyms6q6i22gyam
Differential Revision: https://secure.phabricator.com/D21432
Summary:
Reverts D21419. See PHI1814. Previously, I used "user-select: all" to group sequences of spaces for selection.
However, this has a side effect: the sequence is now selected with a single click. I didn't read the docuementation on the CSS property thoroughly and missed this in testing, since I was focused on drag-selection behavior.
This behavior is enough of a net negative that I think we're in a worse state overall; revert it.
Test Plan: Straight revert.
Differential Revision: https://secure.phabricator.com/D21429
Summary:
Ref T13556. These options are very old and effectively obsoleted by "bin/phd debug [--trace]". I haven't used either option diagnostically in many years, and they aren't mentioned in the documentation.
Remove them to simplify configuration, and because "phd.trace" doesn't work anyway and likely hasn't for a long time -- it has specific issues with TTY detection (see T13556).
Test Plan: Grepped for "phd.trace" and "phd.verbose". Ran "bin/phd debug [--trace]" and saw verbose/trace output.
Maniphest Tasks: T13556
Differential Revision: https://secure.phabricator.com/D21426
Summary:
Ref T13555. Currently, the daemon future may resolve into a failure state immediately inside "start()", and not have a valid PID when we read it.
Instead, read PIDs from the current active future in all cases, using "hasPID()" to test for the presence of a valid PID.
Since we don't query the PID immediately, we no longer need to explicitly start the future.
Also fix an issue where the same future could be added to the overseer pool more than once if it threw on "resolve()". In general:
- Before we "resolve()" a future, detach it from the DaemonHandle: we're always done with it.
- Catch exceptions on resolution and treat them the same way as subprocess resolution errors. These aren't common, but are possible in the general case.
- Have DaemonHandle add futures to the future pool directly when they're created.
Test Plan:
- Ran daemons with intentional subprocess creation failures, saw clean recovery.
- Ran daemons with intentional resolution exceptions, saw clean recovery.
Maniphest Tasks: T13555
Differential Revision: https://secure.phabricator.com/D21425
Summary:
Ref T13555. Although these callsites may not actually impact anything, it's possible for an active handle to have no PID (e.g., if the subprocess failed to start).
Handle these cases more carefully.
Test Plan: Started daemons, saw them run fine. See also next change.
Maniphest Tasks: T13555
Differential Revision: https://secure.phabricator.com/D21424
Summary:
Fixes T13554. For certain prose diff inputs and PCRE backtracking limits, this regular expression may back track too often and fail.
A characteristic input is "x x x x ...", i.e. many sequences where `(.*?)\s*\z` looks like it may be able to match but actually can not.
I think writing an expression which has all the behavior we'd like without this backtracking issue isn't trivial (at least, I don't think I know how to do it offhand); just use a strategy based on "trim()" insetad, which avoids any PCRE complexities here.
Test Plan: Locally, this passes the "x x x ..." test which the previous code failed. I'm not including that test because it won't reproduce across values of "pcre.backtrac_limit", PCRE versions, etc.
Maniphest Tasks: T13554
Differential Revision: https://secure.phabricator.com/D21422
Summary: See PHI1819. This structure may have `null` elements.
Test Plan: Will confirm user reproduction case.
Differential Revision: https://secure.phabricator.com/D21420
Summary:
Ref T2495. See PHI1814. Currently, Phabricator replaces tabs with spaces when rendering diffs.
This may or may not be the best behavior in the long term, but it gives us more control over expansion of tabs than using tab literals.
However, one downside is that you can use your mouse cursor to select "half a tab", and can't use your mouse cursor to distinguish between tabs and spaces. Although you probably shouldn't be doing this, this behavior is less accurate/correct than selecting the entire block as a single unit.
A specific correctness issue with this behavior is that the entire block is copied to the clipboard as a tab literal if you select any of it, so two different visual selection ranges can produce the same clipboard content.
This particular behavior can be improved with "user-select: all", to instruct browsers to select the entire element as a single logical element. Now, selecting part of the tab selects the whole thing, as though it were really a tab literal.
(Some future change might abandon this approach and opt to use real tab literals with "tab-size" CSS, but we lose some ability to control alignment behavior if we do that and it doesn't have any obvious advantages over this approach other than cursor selection behavior.)
Test Plan:
- In Safari and Firefox, dragged text to select a whitespace-expanded tab literal. Saw browsers select the whole sequence as though it were a single tab.
- In Chorme, this also mostly works, but there's some glitchiness and flickering. I think this is still a net improvement, it's just not as smooth as Safari and Firefox.
Maniphest Tasks: T2495
Differential Revision: https://secure.phabricator.com/D21419
Summary:
See PHI1810. In situations where:
- An author submits an urgent change for review.
- The author pings reviewers to ask them to look at it.
...the reviewers may not be able to move the review forward if the review is currently a "Draft". They can only "Commandeer" or ask the author to "Request Review" as ways forward.
Although I'm hesitant to support review actions (particularly, "Accept") on draft revisions, I think there's no harm in allowing reviewers to skip tests and promote the revision out of draft as an explicit action.
Additionally, lightly specialize some of the transaction strings to distinguish between "request review from draft" and other state transitions.
Test Plan:
- As an author, used "Request Review" to promote a draft and to return a change to reviewers for consideration. These behaviors are unchanged, except "promote a draft" has different timeline text.
- As a non-author, used "Begin Review" to promote a draft.
Differential Revision: https://secure.phabricator.com/D21403
Summary:
Currently, adding subscribers to a draft revision raises a warning that they won't get an email/notification.
This warning has some false positives:
- it triggers on any subscriber change, including removing subscribers; and
- it triggers if you're only adding yourself as a subscriber.
Narrow the scope of the warning so it is raised only if you're adding a subscriber other than yourself.
Test Plan:
- Added a non-self subscriber, got the warning as before.
- Added self as a subscriber, no warning (previously: warning).
- Removed a subscriber, no warning (previously: warning).
Differential Revision: https://secure.phabricator.com/D21402
Summary:
See PHI1810. Build toward support for "Request Review" by non-authors on drafts, to forcefully pull a revision out of draft.
Currently, some action strings can't vary based on revision state or the current viewer, so this "pull out of draft" action would have to either: say "Request Review"; or be a totally separate action.
Neither seem great, so allow the labels and messages to vary based on the viewer and revision state.
Test Plan: Grepped for affected symbols, see followup changes.
Differential Revision: https://secure.phabricator.com/D21401
Summary:
Modern Mercurial may emit some more patterns under "--debug".
This whole list is gross and can likely now be eliminated by increasing the minimum required Mercurial version (as `arc` has), but just paper over it for now.
Test Plan:
Locally, saw some views return to functional behavior that weren't previously working on a modern version of Mercurial.
The reproduction case is likely something in the vein of "repository is not writable by webserver, look at history view".
Differential Revision: https://secure.phabricator.com/D21398
Summary:
See PHI1809. This query may join the "slug" table, but each project may have multiple slugs, and the query does not "GROUP BY" when this join occurs.
This may lead to partial result sets and unusual paging behavior.
This could likely be caught categorically in `loadAllFromArray()`; I'll adjust this in a followup.
Test Plan:
A minimal reproduction case is something like:
- Give project P slugs: a, b, c.
- Give project Q slugs: d.
- Query for slugs: a, b, c, d; with limit 2.
- Order the query so P returns first.
- Expect: P and Q.
- Actual: P generates 3 raw rows and the final result is just P with no pagination cursor.
Differential Revision: https://secure.phabricator.com/D21399
Summary:
A handful of Phacility production shards have run into memory pressure issues recently. Although there's no smoking gun, and at least two other plausible contributors, one possible concern is that the Fact daemon was written before hibernation and can not currently hibernate. Even if there's no memory leak, this creates unnecessary memory pressure by holding the processes in memory.
Allow the Fact daemon to hibernate, like other daemons do.
Test Plan: Ran "bin/phd debug fact", saw the Fact daemon hibernate.
Subscribers: yelirekim
Differential Revision: https://secure.phabricator.com/D21389
Summary: Ref T13546. Companion change to D21372. Move URI normalization code to Arcanist to we can more-often resolve remote URIs correctly.
Test Plan: Grepped for affected symbols.
Maniphest Tasks: T13546
Differential Revision: https://secure.phabricator.com/D21373
Summary:
Ref PHI1798. If you put an SSH public key in a table cell with monospaced formatting and then print the table, the cell scrolls and not all of the content appears in your physical printed document.
Generally, the current scrolling behavior for monospaced text seems never-desirable: I can't imagine any cases where we want the table cell to scroll. (There's more of an argument for complex cases where a table cell has, say, an embedded paste.)
Add `line-break: anywhere` to break monospaced text inside these cells.
Test Plan: In Safari, Firefox, and Chrome, viewed a ##|`MMMMM....`|## table. Saw scrolling before and wrapping/breaking after.
Differential Revision: https://secure.phabricator.com/D21370
Summary:
See PHI1794, which describes a connection exhaustion issue with a large number of webhook tasks in queue.
The "GlobalLock" mechanism manages a separate connection pool from the main pool, and webhook workers immediately try to grab a webhook lock with a 0-second wait when they start. So far, this is fine.
Prior to this change, good connections which fail to acqiure a lock are discarded. This can lead to connection exhaustion as the worker rapidly cycles through lock attempts: the connections will remain open for at least 60 seconds (since D16389) in an effort to avoid outbound port exhaustion, but they're effectively orphaned because they aren't part of the main pool and aren't part of the lock pool. We're basically leaking a connection every time we fail to lock.
Failing to lock doesn't mean we need to discard the connection: it's a completely suitable connection for reuse. Instead of dropping it on the floor, put it into the lock pool.
Test Plan:
- Used "bin/webhook call ... --count 10000 --background" to queue a large number of webhook calls against a slow ("sleep(15);") webhook.
- Used "bin/phd launch 32 taskmaster" to start taskmasters.
- Observed MySQL connection behavior:
- Before change: 2048 configured connections immediately exhausted.
- After change: connections stable at ~160ish.
- Ran queue for a while, saw expected single-threaded calls to webhook.
Differential Revision: https://secure.phabricator.com/D21369
Summary:
See PHI1794, which reports an issue where a large number of queued webhook calls led to connection exhaustion. To make this easier to reproduce and test, add "--count" and "--background" flags to "bin/webhook call".
This primarily supports "bin/webook call ... --background --count 10000" to quickly fill the queue with a bunch of calls.
Test Plan: Ran `bin/webhook call` in foreground and background modes, with and without counts. Saw appropriate console and queue behavior.
Differential Revision: https://secure.phabricator.com/D21368
Summary: See PHI1784. Currently, users who pass an invalid SSH command to Phabricator's SSH handler get an unhelpful error message. Make it more helpful.
Test Plan: Ran `./bin/ssh-exec` with no arguments (old, helpful error), invalid arguments (before: unhelpful error; after: helpful error), and valid arguments (old, helpful behavior).
Differential Revision: https://secure.phabricator.com/D21362
Summary:
The "Export Data" workflow incorrectly uses the "Policy Favorites" setting to choose a default export format. This is just a copy/paste error; the correct setting exists and is unused.
If the setting value is an array (as the "Policy Favorites" value often is), we try to use it as an array index. This generates a runtime exception after D21044.
```
[2020-06-16 06:32:12] EXCEPTION: (RuntimeException) Illegal offset type in isset or empty at [<arcanist>/src/error/PhutilErrorHandler.php:263]
#0 <#2> PhutilErrorHandler::handleError(integer, string, string, integer, array) called at [<phabricator>/src/applications/search/controller/PhabricatorApplicationSearchController.php:460]
```
- Use the correct setting.
- Make sure the value we read is a string.
Test Plan:
- Used "Export Data" with a nonempty, array-valued "Policy Favorites" setting.
- Before: runtime exception.
- After: clean export.
- Used "Export Data" again, saw my selection from the first time persisted.
Differential Revision: https://secure.phabricator.com/D21361
Summary: Ref T13546. This makes some "arc" tasks a little easier, and will make them more correct if "arc" ever switches to using SSH.
Test Plan: Ran "harbormaster.buildable.search" from the web UI, saw URIs in the result set.
Subscribers: PHID-OPKG-gm6ozazyms6q6i22gyam
Maniphest Tasks: T13546
Differential Revision: https://secure.phabricator.com/D21346
Summary:
See <https://discourse.phabricator-community.org/t/bad-regex-in-prose-diff-logic/3969>.
The prose splitting rules normally guarantee that newlines appear only at the beginning or end of blocks. However, if a prose sentence ends with text like "...x\n.", we can end up with a newline inside a "sentence".
If we do, the regular expression that breaks it into pieces will fail.
Arguably, this is an error in how sentences are split apart (we might prefer to split this into two sentences, "x\n" and ".", rather than a single "x\n." sentence) but in the general case it's not unreasonable for blocks to contain newlines, so a simple fix is to make the pattern more robust.
Test Plan: Added a failing test which includes this behavior, made it pass.
Differential Revision: https://secure.phabricator.com/D21295
Summary:
Currently, Phortune attempts to prevent users from removing themselves as account managers. It does this by checking that the new list includes them.
Usually this is sufficient, because you can't normally edit an account unless you're already a manager. However, we get the wrong result (incorrect rejection of the edit) if the actor is omnipotent and the acting user was not already a member.
It's okay to edit an account into a state which doesn't include you if you have permission to edit the account and aren't already a manager.
Specifically, this supports more formal tooling around staff modifications to billing accounts, where the actor has staff-omnipotence and the acting user is a staff member and only used for purposes of leaving a useful audit trail.
Test Plan: Elsewhere, ran staff tooling to modify accounts and was able to act as "alice" to add "bailey", even though "alice" was not herself a manager.
Differential Revision: https://secure.phabricator.com/D21288
Summary:
Ref T13513. An inline is not considered empty if it has a suggestion, but some of the shared transaction code doesn't test for this properly.
Update the shared transaction code to be aware that application comments may have more complex emptiness rules.
Test Plan:
- Posted an inline with only an edit suggestion, comment went through.
- Tried to post a normal empty comment, got an appropriate warning.
Maniphest Tasks: T13513
Differential Revision: https://secure.phabricator.com/D21287
Summary: See PHI1753. This condition got rewritten for suggested edits and accidentally inverted.
Test Plan:
- Create a comment, type text, save draft, edit comment, cancel.
- Before: comment hides itself.
- After: comment properly cancels into pre-edit draft state.
Differential Revision: https://secure.phabricator.com/D21286
Summary:
Ref T13541. The passthru future does not have time limit behavior, so if we reach this code we currently fail.
Phabricator never reaches this code normally, but this code is reachable during debugging if you try to foreground a slow fetch to inspect it.
Passthru commands generally only make sense to run interactively, and the caller or control script can enforce their own timeouts (usually by pressing "^C" with their fingers).
Test Plan: Used a debugging script to run ref-by-ref fetches in the foreground.
Maniphest Tasks: T13541
Differential Revision: https://secure.phabricator.com/D21284
Summary:
See PHI1752.
- Early exit of document layout can cause us to fail to populate available rows.
- Some Jupyter documents have "markdown" cells with plain strings, apparently.
Test Plan: Successfully rendered example diff from PHI1752.
Differential Revision: https://secure.phabricator.com/D21285
Summary: Ref T13513. When a revision has inlines with edit suggestions, pressing "j" and "k" can incorrectly select the blocks inside the diffs inside the inlines.
Test Plan: Used "j" to cycle through changes in a revision with inline comments with edit suggestions, didn't get jumped into the suggestion diffs.
Maniphest Tasks: T13513
Differential Revision: https://secure.phabricator.com/D21283
Summary:
Ref PHI1749. Instead of opening files to the last unchanged line on either side of the change, open files to the "simple" line number of the selected block.
For inlines, this is the inline line number.
For blocks, this is the first new-file line number, or the first old-file line number if no new-file line number exists in the block.
This may not always be what the user is hoping for (we can't know what the state of their working copy is) but should produce more obvious behavior.
Test Plan:
- In Diffusion, used "Open in Editor" with and without line selections. Saw same behavior as before.
- Used "n" and "r" to leave an inline with the keyboard, saw same behavior as before.
- Used "\" and "Open in Editor" menu item to open a file with:
- Nothing selected or changeset selected (line: 1).
- An inline selected (line: inline line).
- A block selected (line: first line in block, per above).
Differential Revision: https://secure.phabricator.com/D21282
Summary: Ref T13276. Ref T13513. All readers and writers were removed more than a year ago; clean up the last remnants of this table.
Test Plan: Grepped for table references, found none.
Subscribers: PHID-OPKG-gm6ozazyms6q6i22gyam
Maniphest Tasks: T13513, T13276
Differential Revision: https://secure.phabricator.com/D21281
Summary:
Ref T13513. Syntax highlighting is potentially expensive, and the changeset rendering pipeline can cache it. However, the cache is currently keyed ONLY by Differential changeset ID.
Destroy the existing cache and rebuild it with a more standard cache key so it can be used in a more ad-hoc way by inline suggestion snippets.
Test Plan: Used Darkconsole, saw cache hits and no more inline syntax highlighting for changesets with many inlines.
Subscribers: PHID-OPKG-gm6ozazyms6q6i22gyam
Maniphest Tasks: T13513
Differential Revision: https://secure.phabricator.com/D21280
Summary: Ref T13513. Inline comment context information is somewhat expensive to construct and can be cached. Add a readthrough cache on top of it.
Test Plan: Loaded a source code changeset with many inline comments, used Darkconsole to inspect query activity. Saw caches get populated. Updated cache key, saw caches regenerate. Browsed Diffusion, nothing looked broken.
Maniphest Tasks: T13513
Differential Revision: https://secure.phabricator.com/D21279
Summary: Ref T13513. For now, I'm not supporting inline edit suggestions in Diffusion, although it's likely not difficult to do so in the future. Clean up some of the code so that plain ol' inlines work the same way they did before.
Test Plan:
- Created, edited, reloaded, submitted inlines in Diffusion: familiar old behavior.
- Created, edited, reloaded, submitted inlines with suggestions in Differential: familiar new behavior.
Maniphest Tasks: T13513
Differential Revision: https://secure.phabricator.com/D21278
Summary: Ref T13513. When rendering an inline suggestion for display, use highlighting and diffing.
Test Plan: {F7495053}
Maniphest Tasks: T13513
Differential Revision: https://secure.phabricator.com/D21277
Summary:
Ref T13513. This still has quite a few rough edges and some significant performance isssues, but appears to mostly work.
Allow reviewers to "Suggest Edit" on an inline comment and provide replacement text for the highlighted source.
Test Plan: Created, edited, reloaded, and submitted inline comments in various states with and without suggestion text.
Maniphest Tasks: T13513
Differential Revision: https://secure.phabricator.com/D21276
