Clarifications to network bug triage process.

net/docs/bug-triage-suggested-workflow.txt

+Look for new crashers:
+* Go to go/chromecrash.
+* Choose the platforms and releases for which you will be
+  investigating crashers.  As per bug-triage.txt, this should be the
+  most recent canary, the previous canary (if the most recent is less
+  than a day old), and any of dev/beta/stable that were released in the
+  last couple of days.  These releases should be investigated for all
+  platforms. 
+* For each release, in the "Process Type" frame, click on "browser".
+* At the bottom of the "Magic Signature" frame,  click "limit 1000".
+  Reported crashers are sorted in decreasing order of the number of reports for
+  that crash signature.
+* Search the page for "net::".  
+* For each found signature:
+  * If there is a bug already filed, make sure it is correctly
+    describing the current bug (e.g. not closed, or not describing a
+    long-past issue).  If there is a valid associated bug, ignore this
+    signature. 
+  * Ignore signatures that only occur once, as memory corruption can
+    easily cause one-off failures when the sample size is large
+    enough.
+  * Ignore signatures that only come from a single client ID, as
+    individual machine malware and breakage can also easily cause
+    one-off failures.  
+  * Click on the number of reports field to see details of
+    crash. Ignore it if it doesn't appear to be a network bug. 
+  * Otherwise, file a new bug directly from chromecrash.  Note that
+    this may result in filing bugs for low- and very-low- frequency
+    crashes.  That's ok; the bug tracker is a better tool to figure
+    out whether or not we put resources into those crashes than a snap
+    judgement when filing bugs.

[Ryan Sleevi, 2015/03/02 19:01:09]

There is more needed here - for example, examining third-party modules for signs
of malware (known and unknown)

More importantly, however, the TPMs are already executing this crasher triage as
part of their TPM duties, and already providing a high quality signal/noise
filtering. That is, considering population of distribution to population of
crashes.

[mmenke, 2015/03/02 19:15:13]

I'd take issue with "high quality signal/noise filtering".  I've repeatedly seen
something called a "top-10 crasher" when only one user had 3 or 4 crashes in a
couple hours of Canary data.

I really don't feel like we're doing a good or consistent job of catching
crashers.  We're doing a worse job of investigating crashers, but that's beyond
the scope of this document.  :(

[Randy Smith (Not in Mondays), 2015/03/03 15:08:01]

So I thought about doing that, but I have no clue how to actually write that
section.  Want to suggest some text/pop up your own CL?
Conversation with TPMs in progress; I'll reflect the results in this CL.

[Randy Smith (Not in Mondays), 2015/03/03 15:08:02]

Acknowledged.

+* For each bug you file, include the following information:
+  * The backtrace

[Ryan Sleevi, 2015/03/02 19:01:09]

Do not do this unless you've confirmed with the TPMs on the privacy
implications. This _used_ to be practice and was stopped.

[Randy Smith (Not in Mondays), 2015/03/03 15:08:01]

Conversation in process.

[Randy Smith (Not in Mondays), 2015/03/04 22:22:21]

Refined contexts under which adding a backtrace is appropriate.

+  * The channel in which the bug is seen (canary/dev/beta/stable) and
+    its frequency in that channel.
+  * The frequency of this signature in recent releases.  This
+    information is available by:
+    * Clicking on the signature in the "Magic Signature" list
+    * Clicking "Edit" on the dremel query at the top of the page
+    * Removing the "product.version='X.Y.Z.W' AND" string and clicking
+      "Update".
+    * Clicking "Limit 1000" in the Product Version list in the
+      resulting page (without this, the listing will be restricted to
+      the releases in which the signature is most common, which will
+      often not include the canary/dev release being investigated).
+    * Choose some subset of that list, or all of it, to include in the
+      bug.  Make sure to indicate if there is a defined point in the
+      past before which the signature is not present.
+

[Ryan Sleevi, 2015/03/02 19:01:09]

Please run these last two steps by the TPMs.

[Randy Smith (Not in Mondays), 2015/03/03 15:08:02]

These steps are ok by them (laforge@ specifically).

 Identifying unlabeled network bugs on the tracker:
 * Look at new uncomfirmed bugs since noon PST on the last triager's rotation:
     https://code.google.com/p/chromium/issues/list?can=2&q=status%3Aunconfirmed&sort=-id&num=1000
 * Press "h" to bring up a preview of the bug text.
 * Use "j" and "k" to advance through bugs.
 * If a bug looks like it might be network/download/safe-browsing related, middle
     click [or command-click on OSX] to open in new tab.
 * If a user provides a crash ID for a crasher for a bug that could be
     net-related, look at the crash stack at go/crash, and see if it looks to be
     network related.  Be sure to check if other bug reports have that stack
     trace, and mark as a dupe if so.  Even if the bug isn't network related,
     paste the stack trace in the bug, so no one else has to look up the crash
     stack from the ID.
   * If there's no other information than the crash ID, ask for more details and
       add the Needs-Feedback label.
 * If network causes are possible, ask for a net-internals log (If it's not a
     browser crash) and attach the most specific internals-network label that's
     applicable.  If there isn't an applicable narrower label, a clear owner for
     the issue, or there are multiple possibilities, attach the internals-network
     label and proceed with further investigation.
 * If non-network causes also seem possible, attach those labels as well.
 
+Request data about recent unassigned Cr-Internals-Network bugs from reporters. 
+* For all bugs in this category marked "Needs-Feedback", go through
+  them and remove that label if the reporter has responded with the
+  requested information.
+* For all bugs in this category not marked "Needs-Feedback", go
+  through those bugs and determine if more information is needed from
+  the reporter to make forward progress on the bug.  IF it is, request

[mmenke, 2015/03/02 19:15:13]

nit:  IF -> If

[Randy Smith (Not in Mondays), 2015/03/03 15:08:02]

Done.

+  that information and mark the bug Needs-Feedback.  
+
 Investigating Cr-Internals-Network bugs:
 * It's recommended that while on triage duty, you subscribe to the
     Cr-Internals-Network label.  To do this, go to
     https://code.google.com/p/chromium/issues/ and click on "Subscriptions".
     Enter Cr-Internals-Network and click submit.
 * Look through uncomfirmed and untriaged Cr-Internals-Network bugs, prioritizing
     those updated within the last week:
     https://code.google.com/p/chromium/issues/list?can=2&q=Cr%3DInternals-Network+-status%3AAssigned+-status%3AStarted+-status%3AAvailable+&sort=-modified
 * While investigating a new issue, change the status to Untriaged.
 * If a bug is a potential security issue (Allows for code execution from remote
@@ skipping 35 matching lines @@
     strongly suspect CLs.
 * If you are having trouble with an issue, particularly for help understanding
     net-internals logs, email the public net-dev@chromium.org list for help
     debugging.  If it's a crasher, or for some other reason discussion needs to
     be done in private, use chrome-network-debugging@google.com.
     TODO(mmenke): Write up a net-internals tips and tricks docs.
 * If it appears to be a bug in the unowned core of the network stack (i.e. no
     sublabel applies, or only the Cr-Internals-Network-HTTP sublabel applies,
     and there's no clear owner), try to figure out the exact cause.
 
-Look for new crashers:
-* Go to go/chromecrash.
-* For each platform, go to the latest canary.
-* In the "Process Type" frame, click on "browser".
-* At the bottom of the "Magic Signature" frame,  click "limit 1000".
-  Reported crashers are sorted in decreasing order of the number of reports for
-  that crash signature.
-* Search the page for "net::".  Ignore crashes that only occur once, as
-    memory corruption can easily cause one-off failures when the sample size is
-    large enough.
-* Click on the number of reports field to see details of crash. Look at the
-    stack trace to confirm it's a network bug.
-  * If it is, and there's no associated bug filed, file a new bug directly from
-      chromecrash, looking at earlier canaries to determine if it's a recent
-      regression.  Use the most specific label possible.
-* The most recent Canary may not yet have a full day of crashes, so it may be
-    worth looking at more than one version.
-* If there's been a dev, beta, or stable release in the last couple days, should
-    also look at those.
+Monitor UMA histograms and gasper alerts.  For each Gasper alert that
+fires, determine if it's a real alert and file a bug if so.  
+* Don't file if the alert is coincident with a major volume change.
+  The volume at a particular date can be determined by hovering the
+  mouse over the appropriate location on the alert line.
+* Don't file if the alert is on a graph with very low volume (< ~200
+  data points); it's probably noise, and we probably don't care even
+  if it isn't.
+* Don't file if the graph is really noisy (but eyeball it to decide if
+  there is an underlying important shift under the noise).
+* Don't file if the alert is in the "Known Ignorable" list:
+  * SimpleCache on Windows
+  * DiskCache on Android.
+For each Gasper alert, respond to chrome-network-debugging@ with a
+summary of the action you've taken and why, including issue link if an
+issue was filed.  
 
 Investigating crashers:
 * Only investigate crashers that are still occurring, as identified by above
     section.  If a search on go/crash indicates a crasher is no longer
     occurring, mark it as WontFix.
 * Particularly for Windows, look for weird dlls associated with the crashes.
     If there are some, it may be caused by malware.  You can often figure out if
     a dll is malware by a search, though it's harder to figure out if a dll is
     definitively not malware.
 * See if the same users are repeatedly running into the same issue.  This can be
@@ skipping 17 matching lines @@
   * If a bug is over 2 months old, and the underlying problem was never
     reproduced or really understood:
     * If it's over a year old, go ahead and mark the issue as Archived.
     * Otherwise, ask reporters if the issue is still present, and attach the
         Needs-Feedback label.
 * Old unconfirmed or untriaged Cr-Internals-Network issues can be investigated
     just like newer ones.  Crashers should generally be given higher priority,
     since we can verify if they still occur, and then newer issues, as they're
     more likely to still be present, and more likely to have a still responsive
     bug reporter.