[Esip-documentation] Let's get rid of spatial and temporal bounds in ACDD
Ted Habermann
thabermann at hdfgroup.org
Tue Mar 18 14:22:14 EDT 2014
Would be great!
On Mar 18, 2014, at 11:57 AM, John Graybeal <jbgraybeal at mindspring.com<mailto:jbgraybeal at mindspring.com>> wrote:
Ted, all, any objection to creating a tracking page on the ESIP site (could be with the ACDD pages, or not) describing the known software, point of contact, how it behaves, and (eventually) proposed change(s)?
John
On Mar 18, 2014, at 10:42, Ted Habermann <thabermann at hdfgroup.org<mailto:thabermann at hdfgroup.org>> wrote:
All,
Just wanted to point out that CF is really use metadata used by tools that are actually reading the data. ACDD are discovery conventions originally motivated by the lack of discovery information included in CF. The opinion of the CF mailing list is, therefore, not relevant in this discussion.
I agree with Ken and Ed... There are two sets of metadata because they serve two communities. We need to do our best to make sure they are both correct.
Seems like the netCDF download service in THREDDS is a good place to start... Does anyone know how it behaves with global attributes? What about LAS?
Ted
On Mar 18, 2014, at 10:52 AM, Armstrong, Edward M (398M) <Edward.M.Armstrong at jpl.nasa.gov<mailto:Edward.M.Armstrong at jpl.nasa.gov>> wrote:
Hi Steve,
I do agree with Ken and Ted on this issue….also from experience working with granules in a large data center.
I think the crux of the counter argument is that from the perspective of the data producer we want them to add more metadata, not less to the native granule. And many people do use these global bounds in some context, mostly just looking at the granules through a browser like ncdump and Panoply. Space and time bounds is just a natural thing most people look at in a granule when they first acquire it.
The key, then is to recognize if the granule has been altered by a tool, to treat the bounds with caution. Nan had a great idea that tools should point back to the original unaltered granule. I think this is first requirement for any tool that does modification to the granule via subsetting or similar.
And the next step is to encourage the updating of global bounds by any tool or aggregation operations. I do agree that is a challenge.
On Mar 18, 2014, at 8:51 AM, Steve Hankin <steven.c.hankin at noaa.gov<mailto:steven.c.hankin at noaa.gov>> wrote:
Hi Ken,
I think you are actually agreeing with me, rather than Ted. The text that I proposed did not say geo-bounds attributes should be forbidden. It said that they should be used with caution -- only in situations where they are unlikely to lead to corruption.
It is recommended that 1) the use of these global attributes be restricted to files whose contents are known to be completely stable -- i.e. files very unlikely to be aggregated into larger collections; and 2) as a matter of best practice, software reading CF files should ignore these global attributes; instead it should compute the geo-spatial bounds by scanning the coordinate ranges found within the CF dataset, itself.
These words can certainly be adjusted and improved. (The discussion that I am still hoping will happen!) The intent is to apply common sense, practical thinking so that our interoperability frameworks works in practice.
You have argued that at NODC you *need* these geo-bounds attribute in order to avoid impossibly large processing burdens of examining the data, itself. What would your strategy be if the attributes proved not to be trustworthy? Perhaps within the community of satellite swath folks you can agree that you will all maintain these attributes faithfully. Great. Do so. That is in the spirit of the discussion that we should be having. But should you speak for the modeling community, too? (I presume the NODC granules must be swath data, because if they were grids, then the processing required to determine the bounds from the data is negligible.)
Lets look at the context of this discussion. What we are seeing here is a collision between the priorities of two communities involved in developing standards. Our emails here are largely confined to one of the communities. Envision taking this discussion topic to the CF email list. I predict what you would see is that the potential of redundant, static global attributes to create corruption to the datasets would fly up as a giant red flag. It is not only aggregation operations that will lead to corruption, it is also subsetting operations -- the most routine of all netCDF operations, and performed by generic utilities that are unaware of CF or ACDD. These are glaring, unsolved problems in the use of these geo-bounds. There is such a strong case to be more nuanced when standardizing attributes that we can plainly see are going to lead to many corrupted datasets. What is the counter-argument for ignoring these self-evident problems?
- Steve
===========================================
On 3/18/2014 3:48 AM, Kenneth S. Casey - NOAA Federal wrote:
Hi All -
I've been silent but following this thread carefully. Time to jump in now.
I concur with Ted's statements below. I would characterize his responses as a good example of combating the "Tyranny of the Or". It doesn't have to be one solution or the other. For us at NODC, processing literally tens and maybe hundreds of millions of netCDF granules, having easily and quickly read global attributes is not a convenience. It is a practical necessity. We also like the idea of encouraging softwarians to write better software. Corrupting attributes through negligence and inaction is not acceptable. Fix it. If they can't, we can encourage our users to stop using that software.
I love the idea of building the congruence checker into the ACDD rubric and catalog cleaner, and I think an "ncdump -bounds" option, where the result is calculated from the actual bounds, is terrific too. These kinds of additions to exiting tools would help encourage better practices and would give us some simple tools to improve our management of netCDF data.
Ken
On Mar 17, 2014, at 4:35 PM, Ted Habermann <thabermann at hdfgroup.org<mailto:thabermann at hdfgroup.org>> wrote:
All,
This discussion is driving me bananas!
I would argue (vehemently) against any recommendation that says anything like "don't add bounds elements to your global attributes". To encourage people to provide less metadata is just not acceptable. Much more palatable to
1) offer a general warning to users - all data and metadata have varying quality - test before you leap and know how the data you have was processed.
2) to start a concerted effort to encourage software developers to think about the metadata in the files they create and
3) add a congruence check into the ACDD rubric / catalog cleaner, it would be very interesting to know how many datasets in the "clean" catalog suffer from inconsistencies between the data and the global attributes.
As I said earlier, there are much more significant problems in ACDD/CF metadata land than this one. Hopefully they will generate the same amount of interest as we move forward.
Ted
On Mar 17, 2014, at 1:22 PM, Steve Hankin <steven.c.hankin at noaa.gov<mailto:steven.c.hankin at noaa.gov>> wrote:
Hi John, Rich,
The 'bananas' analogy fits right into the discussion below, so I'm not commenting on I specifically. I like it as a good metaphor for the question we are debating.
On 3/17/2014 10:51 AM, John Graybeal wrote:
Steve,
Let me see if you take issue with my simplified version of what you said:
"NetCDF attributes should never be used to describe value-based or processing-specific features of any data set, because data processors (tools and people) can't help but corrupt derivative data sets with that information. "
I like this line of reasoning for understanding the issues, John. But the above is not a sufficiently nuanced statement to capture what I have advocated. Thy this:
"Broad standardization of NetCDF attributes that contain value-based or processing-derived information about a data set, should be avoided -- used only if there is no reasonable alternative. Use of such attributes breaks the 'backwards compatibility' goals that evolving software standards should follow. It 'breaks' existing systems and places a burden upon future systems that may modify or extend the contents of that dataset."
Standardization of such attributes for special communities that run custom systems, presents no serious problems. CF allows itself to be arbitrarily extended for such special purposes. Each community will look after its own software quality. But when you undertake a broad standardization of such attributes, you are placing a burden on software systems over which you have no knowledge and no control.
Don't get me wrong; I think this presents a consistent philosophy in response to today's practical realities. As of today, with 5 out of 6 tools failing the 'metadata consistency' test, it likely minimizes the ratio of bad metadata in the wild.
thank you for this.
What it does *not* do is establish a mature, flexible, interoperable, metadata-aware community of practice going forward. It also does not fix the problem caused by these tools, it just eliminates the most-likely-to-fail attributes (from any standard, forevermore). And it directly undercuts the greatest value of a self-describing data format, the easy-to-access description.
As Ted pointed out, the issue is much broader and deeper than this one particular use case. Dynamically changing datasets (including the 'virtual datasets' created through aggregation) create a class of problems that are in fundamental conflict with static metadata representations. Do we agree that that virtual datasets are not going to go away? The class of datasets that CF is most centrally committed to are often extremely large and evolve over time (both the 'C' and the 'F' in CF).
The fact that much software pre-dates the standards is a red herring.
Really? We watch standards self-destruct time after time because they fail to address pragmatic considerations of this very type. Backwards compatibility deserves to be a paramount consideration. Do we really disagree on this?
Software that takes any self-describing file, modifies the contents, yet passes through the original description _without validation_ to the output file, will almost always produce a broken output file. This isn't a feature, nor inevitable; it's a bug. Our underlying challenge is to fix that problem.
And in the meantime, we still have to maintain ALL the metadata affected by that problem, not just the CF coordinate bounds. So we should fix all that software soon. (As I can think of 3 trivial fixes off the top of my head, I'm not sure why previous beat-arounds didn't induce change. It's time. And it's a lot easier than modifying all the software to include statements in the files about how unreliable the metadata is.)
I agree with you on this ... at least in principle: Those responsible for the software systems should make a good faith effort to upgrade them in response to evolving standards. But they may simply lack the resources. Or they may be unable to fit the fixes into hard-pressed schedules for some time to come. In the meantime we have corrupted metadata. We are living in lean times for our community. Considerations like this have to be a two-way street: those responsible for evolving software standards need to minimize so-called 'enhancements' that break existing software. They need to look for alternatives first.
- Steve
John
On Mar 17, 2014, at 09:23, Steve Hankin <steven.c.hankin at noaa.gov<mailto:steven.c.hankin at noaa.gov>> wrote:
Greetings Ted!
As rants go, the message you sent on Friday was pretty restrained. I particularly like your suggestion that the UAF Catalog Cleaner could detect and report corrupted ACDD geo-positioning attributes. We will see what we can do with that idea. (aside: Is there a way to ask ncISO when it has found this form of corruption in a dataset?)
It would be nice if we could all wrestle this topic to a workable compromise. I agree that the problem should be "fixed at the source". But characterizing the source of the problem as "sloppy data management" seems off the mark. This data management problem didn't exist until we created the potential for easy corruption by defining easily-corrupted, redundant information in the CF datasets.
The duplication of information between the global attributes and the CF coordinates is an 'attractive nuisance'. No matter how much we exhort data managers to clean up their act, the problem is going to continue showing up over and over and over. Some of the data management tools that stumble into this problem are not even CF-aware; generic netCDF utilities like nco create this form of corruption. Offhand I can think of 6 independent pieces of software that perform aggregations on CF datasets. Five of these predate the use of the ACDD geo-bounds attributes within CF files; all of these exhibit this corruption. Ed's example is yet another.
Our underlying challenge is to expose the CF coordinate bounds for easy use wherever that information is needed for purposes of data discovery. The ncISO tool contributed by your group has addressed this very successfully for formal ISO metadata. (A big thanks to you, Dave N. et. al..) There is similar excellent potential to address the more informal cases through software. You mentioned the limitations of "ncdump -h" as an illustration. How about code contributed to Unidata to create "ncdump -bounds"? This would be a smaller effort with a more robust outcome than to ask all current and future developers of CF aggregation techniques to make accommodation for the redundant, easily corrupted attributes in their datasets.
- Steve
===========================================
On 3/14/2014 1:14 PM, Ted Habermann wrote:
All,
I agree with Ed and John, this is a software tool problem that should be fixed at the source. The description of the history attribute has always implied that it should be updated when a file is processed (even though, IMHO, it is almost entirely unsuited for doing that). The same is true for many others (listed by John G. earlier in this thread). The current practice is sloppy data management that, from the sound of this thread, is pervasive in the community. Of course, ncISO provides a very easy way to identify occurrences of this problem throughout a THREDDS catalog. The "Catalog Cleaner" is another venue for quantifying the damage. CF is a community standard. Maybe it is time for the community to recommend providing correct metadata with the files and to avoid developers and datasets that don't.
A related problem is that the bounds calculated from the data are only available if you read the data. Many users may not be equipped to easily read the data during a data discovery process. They may not want to go beyond ncdump -x -h (or something like that) before they fire up the whole netCDF machine...
BTW, this problem is trivial relative to that associated with virtual datasets created through aggregation. In those cases, there is no clear mechanism for providing meaningful metadata, although the rich inventory we created several years ago comes close... That situation is much more prone to mistakes as all semblance of the historic record is wiped out.
Its Friday, and spring... As Dave said last week, a good time for a rant!
Ted
On Mar 14, 2014, at 1:44 PM, Steve Hankin <steven.c.hankin at noaa.gov<mailto:steven.c.hankin at noaa.gov>> wrote:
Hi All,
I'm joining into this discussion from the wings. The topic here -- the common tendency for the ACDD geo-spacio-temporal bounds attributes to get corrupted -- has been beaten around a number of times among different groups. At this point it isn't clear that there is a "clean" resolution to the problem; there are already so many files out there that contain these attributes that there may be no easy way to unwind the problem. Might the best path forward be to see about adding some words of caution into the the documents that suggest the use of these attributes? Something along these lines:
Caution: The encoding of geo-spatial bounds values as global attributes is a practice that should be used with caution or avoided.
The encoding of geo-spatial bounds values as global attributes introduces a high likelihood of corruption, because the attibute values duplicate information already contained in the self-describing coordinates of the dataset. A number of data management operations that are common with netCDF files will invalidate the values stored as global attributes. Such operations include extending the coordinate range of a netCDF file along its record axis; aggregating a collection of netCDF files into a larger datasets (for example aggregating model outputs along their time axes); or appending files using file-based utilities (e.g. nco).
It is recommended that 1) the use of these global attributes be restricted to files whose contents are known to be completely stable -- i.e. files very unlikely to be aggregated into larger collections; and 2) as a matter of best practice, software reading CF files should ignore these global attributes; instead it should compute the geo-spatial bounds by scanning the coordinate ranges found within the CF dataset, itself.
Comments?
- Steve
<http://www.facebook.com/feeds/page.php?id=178512945559611&format=rss20>
<http://www.facebook.com/feeds/page.php?id=178512945559611&format=rss20>
_______________________________________________
Esip-documentation mailing list
Esip-documentation at lists.esipfed.org<mailto:Esip-documentation at lists.esipfed.org>
http://www.lists.esipfed.org/mailman/listinfo/esip-documentation
-ed
Ed Armstrong
JPL Physical Oceanography DAAC
818 519-7607
<SignatureSm.png>
John Graybeal
jbgraybeal at mindspring.com<mailto:jbgraybeal at mindspring.com>
[cid:32323496-C60B-49FF-8310-11CCF46BDC72]
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.lists.esipfed.org/pipermail/esip-documentation/attachments/20140318/603c46e0/attachment-0001.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: SignatureSm.png
Type: image/png
Size: 30402 bytes
Desc: SignatureSm.png
URL: <http://www.lists.esipfed.org/pipermail/esip-documentation/attachments/20140318/603c46e0/attachment-0001.png>
More information about the Esip-documentation
mailing list