There are two heavily related issues here, but I'm bundling them in the same bug.

The primary issue is our use of under/overflow buckets are not very intuitive. In general, we use the 0-1 bucket as an underflow, and bucket_max->int_max as an overflow. One flaw with this is real data can fall in the 0-1 range, and this can get mixed up with incorrect data, in theory. 
For both cases, it would be nice if we supported this a bit better on the dashboard, and gave special labels for real under/overflow data.

A related issue is that this can be confusing to work with in Chromium. Since we mostly use macros, there isn't much of a problem, but our suggestions for inputs to the FactoryGet for LinearHistogram is incorrect.


First, note our suggestion:
https://codesearch.chromium.org/chromium/src/base/metrics/histogram_base.cc?q=LinearHistogram+factoryget&sq=package:chromium&l=143&dr=C
This looks like we want min=1, max=MAX+1, num_buck=NUM_BUCK+1

Here are a couple of bad cases here:
https://codesearch.chromium.org/chromium/src/base/metrics/persistent_memory_allocator.cc?q=persistent+histograms+factoryget+21&sq=package:chromium&l=413&dr=C

Note this is even in base/metrics. The intention clearly was to get a percentage (1-100) with bucket sizes of 5. However, if you look at the data, we see ranges of 5 and 6. The problem is the logic itself assumed both under and overflow buckets includes, 
https://codesearch.chromium.org/chromium/src/base/metrics/histogram.cc?sq=package:chromium&dr=CSs&rcl=1475234468&l=895
(see bucket_count - 2). If we had set it to be 22 and not 21, then we would get ranges of 5, going from 1-101. This is better, but still we would prefer ranges 0-100, and then under/overflow buckets...

Another sample issue:
https://codesearch.chromium.org/chromium/src/content/renderer/pepper/pepper_plugin_instance_metrics.cc?q=LinearHistogram::FactoryGet&sq=package:chromium&l=60&dr=C
Similar, this expected 100 bucket of size 5, but we see a mix of length 5 and 6 in the data.

Inconsistent sizes in linear histogram is pretty bad since (in these examples), a casual viewing of the histogram will show every Nth bucket have a 20% increase in data. Exponential histograms aren't as sensitive but we still have the under/overflow issues there.