Consistent rules for handling merges between variables with different attributes

[shoyer]

Currently, variable attributes are checked for equality before allowing for a merge via a call to xarray_equal. It should be possible to merge datasets even if some of the variable metadata disagrees (conflicting attributes should be dropped). This is already the behavior for global attributes.

The right design of this feature should probably include some optional argument to Dataset.merge indicating how strict we want the merge to be. I can see at least three versions that could be useful:

1. Drop conflicting metadata silently.
2. Don't allow for conflicting values, but drop non-matching keys.
3. Require all keys and values to match.

We can argue about which of these should be the default option. My inclination is to be as flexible as possible by using 1 or 2 in most cases.

[ebrevdo]

I would default to 3, and in the exception suggest using a different merge option. Imagine merging two datasets with different _FillValue, unit, or compression attributes.

[shoyer]

Dataset.merge is also triggered by assigning a DatasetArray to a dataset or by doing a mathematical operation on two DatasetArrays (e.g., x + y). The later is how I encountered this issue today.

For merge itself, I would agree that we may want to default to stricter behavior, but for these other versions of merge we should default to something more flexible.

[ebrevdo]

It depends on whether x+y does attribute checking before performing the
merge. Again, if units don't match then maybe you shouldn't add. I always
favor the strictest approach so you don't get strange surprises.

On Wed, Feb 26, 2014 at 2:47 PM, Stephan Hoyer notifications@github.comwrote:

Dataset.merge is also triggered by assigning a DatasetArray to a dataset
or by doing a mathematical operation on two DatasetArrays (e.g., x + y).
The later is how I encountered this issue today.

For merge itself, I would agree that we may want to default to stricter
behavior, but for these other versions of merge we should default to
something more flexible.

Reply to this email directly or view it on GitHubhttps://github.com/akleeman/xray/issues/25#issuecomment-36187723
.

[shoyer]

x + y could indeed check variable attributes before trying to do the merge. I don't know if it does in the current implementation.

My concern is more that metadata like "title" or "source" should not be required to match, because that metadata will almost always be conflicting. Perhaps "units", "_FIllValue", "scale_factor" and "add_offset" (if values were not automatically masked/scaled) should be specifically blacklisted to prohibit conflicts.

[ebrevdo]

This is an option, but these lists will break if we try to express other
data formats using these conventions. For example, grib likely has other
conventions. We would have to overload attribute or variable depending on
what the underlying datastore is.

On Wed, Feb 26, 2014 at 3:03 PM, Stephan Hoyer notifications@github.comwrote:

x + y could indeed check variable attributes before trying to do the
merge. I don't know if it does in the current implementation.

My concern is more that metadata like "title" or "source" should not be
required to match, because that metadata will almost always be conflicting.
Perhaps "units", "_FIllValue", "scale_factor" and "add_offset" (if values
were not automatically masked/scaled) should be specifically blacklisted to
prohibit conflicts.

Reply to this email directly or view it on GitHubhttps://github.com/akleeman/xray/issues/25#issuecomment-36189171
.

[ebrevdo]

Also, there are plenty of other bits where you don't want conflicts.
Imagine that you have variables indexed on different basemap projections.
Creating exceptions to the rule seems like a bit of a rabbit hole.

On Wed, Feb 26, 2014 at 3:13 PM, Eugene Brevdo ebrevdo@gmail.com wrote:

This is an option, but these lists will break if we try to express other
data formats using these conventions. For example, grib likely has other
conventions. We would have to overload attribute or variable depending on
what the underlying datastore is.

On Wed, Feb 26, 2014 at 3:03 PM, Stephan Hoyer notifications@github.comwrote:

x + y could indeed check variable attributes before trying to do the
merge. I don't know if it does in the current implementation.

My concern is more that metadata like "title" or "source" should not be
required to match, because that metadata will almost always be conflicting.
Perhaps "units", "_FIllValue", "scale_factor" and "add_offset" (if
values were not automatically masked/scaled) should be specifically
blacklisted to prohibit conflicts.

Reply to this email directly or view it on GitHubhttps://github.com/akleeman/xray/issues/25#issuecomment-36189171
.

[shoyer]

I see your point, but I favor a more pragmatic approach by default. See my
fourth bullet under "Design Goals" in the README and bullet ii under Iris
in "Prior Art".

My vision here is a more powerful ndarray enhanced rather than limited by
metadata. This is closer to what pandas does, which even allows for
conflicting indices resulting in NaN values (a feature I would love to
copy).

I think that both use cases can be covered as long as the merge/conflict
logic is clearly documented and it is possible to write stricter logic for
library code (which by necessity will be more verbose). If it is essential
for units to agree before doing x + y, you can add assert x.attribubes.get('units') == y.attributes.get('units'). Otherwise, we will
end up prohibiting operations like that when x has units of Celsius and y
has units of Kelvin.

On Wed, Feb 26, 2014 at 3:23 PM, ebrevdo notifications@github.com wrote:

Also, there are plenty of other bits where you don't want conflicts.
Imagine that you have variables indexed on different basemap projections.
Creating exceptions to the rule seems like a bit of a rabbit hole.

On Wed, Feb 26, 2014 at 3:13 PM, Eugene Brevdo ebrevdo@gmail.com wrote:

This is an option, but these lists will break if we try to express other
data formats using these conventions. For example, grib likely has other
conventions. We would have to overload attribute or variable depending on
what the underlying datastore is.

On Wed, Feb 26, 2014 at 3:03 PM, Stephan Hoyer <notifications@github.com
wrote:

x + y could indeed check variable attributes before trying to do the
merge. I don't know if it does in the current implementation.

My concern is more that metadata like "title" or "source" should not be
required to match, because that metadata will almost always be
conflicting.
Perhaps "units", "_FIllValue", "scale_factor" and "add_offset" (if
values were not automatically masked/scaled) should be specifically
blacklisted to prohibit conflicts.

Reply to this email directly or view it on GitHub<
https://github.com/akleeman/xray/issues/25#issuecomment-36189171>
.

Reply to this email directly or view it on GitHubhttps://github.com/akleeman/xray/issues/25#issuecomment-36190935
.

[ebrevdo]

I don't think that example has your intended affect. I don't know why
anyone would add something of units kelvin with those of celsius. I
understand what you're saying, so maybe we should just throw a stern
warning listing which units conflict and how, every single time.

On Wed, Feb 26, 2014 at 3:42 PM, Stephan Hoyer notifications@github.comwrote:

I see your point, but I favor a more pragmatic approach by default. See my
fourth bullet under "Design Goals" in the README and bullet ii under Iris
in "Prior Art".

My vision here is a more powerful ndarray enhanced rather than limited by
metadata. This is closer to what pandas does, which even allows for
conflicting indices resulting in NaN values (a feature I would love to
copy).

I think that both use cases can be covered as long as the merge/conflict
logic is clearly documented and it is possible to write stricter logic for
library code (which by necessity will be more verbose). If it is essential
for units to agree before doing x + y, you can add assert x.attribubes.get('units') == y.attributes.get('units'). Otherwise, we will
end up prohibiting operations like that when x has units of Celsius and y
has units of Kelvin.

On Wed, Feb 26, 2014 at 3:23 PM, ebrevdo notifications@github.com wrote:

Also, there are plenty of other bits where you don't want conflicts.
Imagine that you have variables indexed on different basemap projections.
Creating exceptions to the rule seems like a bit of a rabbit hole.

On Wed, Feb 26, 2014 at 3:13 PM, Eugene Brevdo ebrevdo@gmail.com
wrote:

This is an option, but these lists will break if we try to express
other
data formats using these conventions. For example, grib likely has
other
conventions. We would have to overload attribute or variable depending
on
what the underlying datastore is.

On Wed, Feb 26, 2014 at 3:03 PM, Stephan Hoyer <
notifications@github.com
wrote:

x + y could indeed check variable attributes before trying to do the
merge. I don't know if it does in the current implementation.

My concern is more that metadata like "title" or "source" should not
be
required to match, because that metadata will almost always be
conflicting.
Perhaps "units", "_FIllValue", "scale_factor" and "add_offset" (if

values were not automatically masked/scaled) should be specifically
blacklisted to prohibit conflicts.

Reply to this email directly or view it on GitHub<
https://github.com/akleeman/xray/issues/25#issuecomment-36189171>
.

Reply to this email directly or view it on GitHub<
https://github.com/akleeman/xray/issues/25#issuecomment-36190935>

.

Reply to this email directly or view it on GitHubhttps://github.com/akleeman/xray/issues/25#issuecomment-36192859
.

[ebrevdo]

err, which attributes conflict.

On Wed, Feb 26, 2014 at 3:45 PM, Eugene Brevdo ebrevdo@gmail.com wrote:

I don't think that example has your intended affect. I don't know why
anyone would add something of units kelvin with those of celsius. I
understand what you're saying, so maybe we should just throw a stern
warning listing which units conflict and how, every single time.

On Wed, Feb 26, 2014 at 3:42 PM, Stephan Hoyer notifications@github.comwrote:

I see your point, but I favor a more pragmatic approach by default. See my
fourth bullet under "Design Goals" in the README and bullet ii under Iris
in "Prior Art".

My vision here is a more powerful ndarray enhanced rather than limited by
metadata. This is closer to what pandas does, which even allows for
conflicting indices resulting in NaN values (a feature I would love to
copy).

I think that both use cases can be covered as long as the merge/conflict
logic is clearly documented and it is possible to write stricter logic for
library code (which by necessity will be more verbose). If it is essential
for units to agree before doing x + y, you can add assert x.attribubes.get('units') == y.attributes.get('units'). Otherwise, we
will
end up prohibiting operations like that when x has units of Celsius and y
has units of Kelvin.

On Wed, Feb 26, 2014 at 3:23 PM, ebrevdo notifications@github.com
wrote:

Also, there are plenty of other bits where you don't want conflicts.
Imagine that you have variables indexed on different basemap
projections.
Creating exceptions to the rule seems like a bit of a rabbit hole.

On Wed, Feb 26, 2014 at 3:13 PM, Eugene Brevdo ebrevdo@gmail.com
wrote:

This is an option, but these lists will break if we try to express
other
data formats using these conventions. For example, grib likely has
other
conventions. We would have to overload attribute or variable
depending on
what the underlying datastore is.

On Wed, Feb 26, 2014 at 3:03 PM, Stephan Hoyer <
notifications@github.com
wrote:

x + y could indeed check variable attributes before trying to do the
merge. I don't know if it does in the current implementation.

My concern is more that metadata like "title" or "source" should not
be
required to match, because that metadata will almost always be
conflicting.
Perhaps "units", "_FIllValue", "scale_factor" and "add_offset" (if

values were not automatically masked/scaled) should be specifically
blacklisted to prohibit conflicts.

Reply to this email directly or view it on GitHub<
https://github.com/akleeman/xray/issues/25#issuecomment-36189171>
.

Reply to this email directly or view it on GitHub<
https://github.com/akleeman/xray/issues/25#issuecomment-36190935>

.

Reply to this email directly or view it on GitHubhttps://github.com/akleeman/xray/issues/25#issuecomment-36192859
.

[shoyer]

Stern warnings about conflicting attributes like units may be an
appropriate compromise. But if we go that way, I would advocate for making
units drop upon doing any mathematical operation. We could try to update
units automatically (e.g., kg * kg = kg^2), but that is tricky to always
right.

Celsius, for example, is a pretty weird physical unit because of how it can
take on negative values, so it actually makes a lot of sense to use mostly
use Kelvin instead (for which I can sensibly apply any math operation like
times or minus). That doesn't mean that I want to store all my raw data in
degrees K, though...

On Wed, Feb 26, 2014 at 3:45 PM, ebrevdo notifications@github.com wrote:

err, which attributes conflict.

On Wed, Feb 26, 2014 at 3:45 PM, Eugene Brevdo ebrevdo@gmail.com wrote:

I don't think that example has your intended affect. I don't know why
anyone would add something of units kelvin with those of celsius. I
understand what you're saying, so maybe we should just throw a stern
warning listing which units conflict and how, every single time.

On Wed, Feb 26, 2014 at 3:42 PM, Stephan Hoyer <notifications@github.com
wrote:

I see your point, but I favor a more pragmatic approach by default. See
my
fourth bullet under "Design Goals" in the README and bullet ii under
Iris
in "Prior Art".

My vision here is a more powerful ndarray enhanced rather than limited
by
metadata. This is closer to what pandas does, which even allows for
conflicting indices resulting in NaN values (a feature I would love to
copy).

I think that both use cases can be covered as long as the merge/conflict
logic is clearly documented and it is possible to write stricter logic
for
library code (which by necessity will be more verbose). If it is
essential
for units to agree before doing x + y, you can add assert x.attribubes.get('units') == y.attributes.get('units'). Otherwise, we
will
end up prohibiting operations like that when x has units of Celsius and
y
has units of Kelvin.

On Wed, Feb 26, 2014 at 3:23 PM, ebrevdo notifications@github.com
wrote:

Also, there are plenty of other bits where you don't want conflicts.
Imagine that you have variables indexed on different basemap
projections.
Creating exceptions to the rule seems like a bit of a rabbit hole.

On Wed, Feb 26, 2014 at 3:13 PM, Eugene Brevdo ebrevdo@gmail.com
wrote:

This is an option, but these lists will break if we try to express
other
data formats using these conventions. For example, grib likely has
other
conventions. We would have to overload attribute or variable
depending on
what the underlying datastore is.

On Wed, Feb 26, 2014 at 3:03 PM, Stephan Hoyer <
notifications@github.com
wrote:

x + y could indeed check variable attributes before trying to do
the
merge. I don't know if it does in the current implementation.

My concern is more that metadata like "title" or "source" should
not
be
required to match, because that metadata will almost always be
conflicting.
Perhaps "units", "_FIllValue", "scale_factor" and "add_offset" (if

values were not automatically masked/scaled) should be specifically
blacklisted to prohibit conflicts.

Reply to this email directly or view it on GitHub<
https://github.com/akleeman/xray/issues/25#issuecomment-36189171>
.

Reply to this email directly or view it on GitHub<
https://github.com/akleeman/xray/issues/25#issuecomment-36190935>

.

Reply to this email directly or view it on GitHub<
https://github.com/akleeman/xray/issues/25#issuecomment-36192859>
.

Reply to this email directly or view it on GitHubhttps://github.com/akleeman/xray/issues/25#issuecomment-36193148
.

[ebrevdo]

Agreed. I would avoid that kind of thing too. Maybe a stern warning for
all conflicting attributes, and saying that they will be dropped from the
new variable.

For units specifically, Python has a variety of unit libraries that wrap
numpy arrays and can probably do some magic. Not sure if we really want to
do that, though.

On Wed, Feb 26, 2014 at 4:07 PM, Stephan Hoyer notifications@github.comwrote:

Stern warnings about conflicting attributes like units may be an
appropriate compromise. But if we go that way, I would advocate for making
units drop upon doing any mathematical operation. We could try to update
units automatically (e.g., kg * kg = kg^2), but that is tricky to always
right.

Celsius, for example, is a pretty weird physical unit because of how it can
take on negative values, so it actually makes a lot of sense to use mostly
use Kelvin instead (for which I can sensibly apply any math operation like
times or minus). That doesn't mean that I want to store all my raw data in
degrees K, though...

On Wed, Feb 26, 2014 at 3:45 PM, ebrevdo notifications@github.com wrote:

err, which attributes conflict.

On Wed, Feb 26, 2014 at 3:45 PM, Eugene Brevdo ebrevdo@gmail.com
wrote:

I don't think that example has your intended affect. I don't know why
anyone would add something of units kelvin with those of celsius. I
understand what you're saying, so maybe we should just throw a stern
warning listing which units conflict and how, every single time.

On Wed, Feb 26, 2014 at 3:42 PM, Stephan Hoyer <
notifications@github.com
wrote:

I see your point, but I favor a more pragmatic approach by default.
See
my
fourth bullet under "Design Goals" in the README and bullet ii under
Iris
in "Prior Art".

My vision here is a more powerful ndarray enhanced rather than limited
by
metadata. This is closer to what pandas does, which even allows for
conflicting indices resulting in NaN values (a feature I would love to
copy).

I think that both use cases can be covered as long as the
merge/conflict
logic is clearly documented and it is possible to write stricter logic
for
library code (which by necessity will be more verbose). If it is
essential
for units to agree before doing x + y, you can add assert x.attribubes.get('units') == y.attributes.get('units'). Otherwise, we
will
end up prohibiting operations like that when x has units of Celsius
and
y
has units of Kelvin.

On Wed, Feb 26, 2014 at 3:23 PM, ebrevdo notifications@github.com
wrote:

Also, there are plenty of other bits where you don't want
conflicts.
Imagine that you have variables indexed on different basemap
projections.
Creating exceptions to the rule seems like a bit of a rabbit hole.

On Wed, Feb 26, 2014 at 3:13 PM, Eugene Brevdo ebrevdo@gmail.com
wrote:

This is an option, but these lists will break if we try to express
other
data formats using these conventions. For example, grib likely has
other
conventions. We would have to overload attribute or variable
depending on
what the underlying datastore is.

On Wed, Feb 26, 2014 at 3:03 PM, Stephan Hoyer <
notifications@github.com
wrote:

x + y could indeed check variable attributes before trying to do
the
merge. I don't know if it does in the current implementation.

My concern is more that metadata like "title" or "source" should
not
be
required to match, because that metadata will almost always be
conflicting.
Perhaps "units", "_FIllValue", "scale_factor" and "add_offset"
(if

values were not automatically masked/scaled) should be
specifically
blacklisted to prohibit conflicts.

Reply to this email directly or view it on GitHub<
https://github.com/akleeman/xray/issues/25#issuecomment-36189171>
.

Reply to this email directly or view it on GitHub<
https://github.com/akleeman/xray/issues/25#issuecomment-36190935>

.

Reply to this email directly or view it on GitHub<
https://github.com/akleeman/xray/issues/25#issuecomment-36192859>
.

Reply to this email directly or view it on GitHub<
https://github.com/akleeman/xray/issues/25#issuecomment-36193148>

.

Reply to this email directly or view it on GitHubhttps://github.com/akleeman/xray/issues/25#issuecomment-36194729
.