OPeNDAP Documentation - User contributions [en]

DMR++

2021-07-26T19:01:12Z

Dan: /* DAP representations */

''How to build & deploy dmr++ files for Hyrax''

==What? Why?==

It is a fast and flexible way to serve data stored in S3.
The dmr++ encodes the location of the data content residing in a binary data file/object (e.g., an hdf5 file) so that it can be directly accessed, without the need for an intermediate library API, by using the file with the location information. The binary data objects may be on a local filesystem, or they may reside across the web in something like an S3 bucket.

==How Does It Work?==

The ''dmr++'' ingest software reads a data file (see '''note''') and builds a document that holds all of the file's metadata (the names and types of all of the variables along with any other information bound to those variables). This information is stored in a document we call the Dataset Metadata Response (DMR). The ''dmr++'' adds some extra information to this (that's the '++' part) about where each variable can be found and how to decode those values. The ''dmr++'' is simply an special annotated DMR document.

This effectively decouples the annotated DMR (''dmr++'') from the location of the granule file itself. Since dmr++ files are typically significantly smaller than the source data granules they represent, they can be stored and moved for less expense. They also enable reading all of the file's metadata in one operation instead of the iterative process that many APIs require.

If the dmr++ contains references to the source granules location on the web, the location of the the dmr++ file itself does not matter.

Software that understands the dmr++ content can directly access the data values held in the source granule file, and it can do so without having to retrieve the entire file and work on it locally, even when the file is stored in a Web Object Store like S3.

If the granule file contains multiple variables and only a subset of them are needed, the dmr++ enabled software can retrieve just the bytes associated with the desired variables values.

'''note:''' The OPeNDAP software currently supports HDF5 and NetCDF4. Other formats can be supported, such as zarr.

==Supported Data Formats==

The ''dmr++'' software currently works with ''hdf5'' and ''netcdf-4'' files. (The ''netcdf-4'' format is a subset of ''hdf5'' so ''hdf5'' tools are utilized for both.) Other formats like ''zarr'', ''hdf4'', ''netcdf-3'' are not currently supported by the ''dmr++'' software, but support could be added if needed.

===''hdf5''===

The ''hdf5'' data format is quite complex and many of the options and edge cases are not currently supported by the ''dmr++'' software.

These limitations and how to quickly evaluate an ''hdf5'' or ''netcdf-4'' file for use with the ''dmr++'' software are explained below.

====''hdf5'' filters====

The ''hdf5'' format has several filter/compression options used for storing data values.
The ''dmr++'' software currently supports data that utilize the H5Z_FILTER_DEFLATE and H5Z_FILTER_SHUFFLE filters.
[https://support.hdfgroup.org/HDF5/doc/RM/RM_H5Z.html You can find more on hdf5 filters here.]

====''hdf5'' storage layouts====

The ''hdf5'' format also uses a number of "storage layouts" that describe various structural organizations of the data values associated with a variable in the granule file.
The ''dmr++'' software currently supports data that utilize the H5D_COMPACT, H5D_CHUNKED, and H5D_CONTIGUOUS storage layouts. These are all of the storage layouts defined by the ''hdf5'' library, but others can be added.
[https://support.hdfgroup.org/HDF5/doc1.6/Datasets.html You can find more on hdf5 storage layouts here.]

====Is my ''hdf5'' or ''netcdf-4'' file suitable for ''dmr++''?====

To determine the ''hdf5'' filters, storage layouts, and chunking scheme used in an ''hdf5'' or ''netcdf-4'' file you can use the command:
<code>h5dump -H -p <filename></code>
To get a human readable assessment of the file that will show the storage layouts, chunking structure, and the filters needed for each variable (aka DATASET in the ''hdf5'' vocabulary) [https://support.hdfgroup.org/HDF5/doc/RM/Tools.html#Tools-Dump h5dump info can be found here.]

''h5dump example output:''
$ h5dump -H -p chunked_gzipped_fourD.h5
HDF5 "chunked_gzipped_fourD.h5" {
GROUP "/" {
DATASET "d_16_gzipped_chunks" {
DATATYPE H5T_IEEE_F32LE
DATASPACE SIMPLE { ( 40, 40, 40, 40 ) / ( 40, 40, 40, 40 ) }
STORAGE_LAYOUT {
CHUNKED ( 20, 20, 20, 20 )
SIZE 2863311 (3.576:1 COMPRESSION)
}
FILTERS {
COMPRESSION DEFLATE { LEVEL 6 }
}
FILLVALUE {
FILL_TIME H5D_FILL_TIME_ALLOC
VALUE H5D_FILL_VALUE_DEFAULT
}
ALLOCATION_TIME {
H5D_ALLOC_TIME_INCR
}
}
}
}

====Is my netcdf file ''netcdf-3'' or ''netcdf-4''?====

It is an unfortunate state of affairs that the file suffix ".nc" is the commonly used naming convention for both ''netcdf-3'' and ''netcdf-4'' files.
You can use the command:
<code>ncdump -k <filename></code>
to determine if a ''netcdf'' file is either classic ''netcdf-3'' (classic) or ''netcdf-4'' (netCDF-4).
* The ''netcdf'' library must be installed on the system upon which the command is issued.

[http://www.bic.mni.mcgill.ca/users/sean/Docs/netcdf/guide.txn_79.html You can learn more in the NetCDF documentation here.]

==Building ''dmr++'' files with get_dmrpp==

The application that builds the ''dmr++'' files is a command line tool called ''get_dmrpp''. It in turn utilizes other executables such as ''build_dmrpp'', ''reduce_mdf'', ''merge_dmrpp'' (which rely in turn on the ''hdf5_handler'' and the ''hdf5'' library), along with a number of UNIX shell commands.

All of these components are install with each recent version of the Hyrax Data Server

You can see the ''get_dmrpp'' usage statement with the command: <code>get_dmrpp -h</code>

===Using ''get_dmrpp''===

The way that ''get_dmrpp'' is invoked controls the way that the data are ultimately represented in the resulting ''dmr++'' file(s).

The ''get_dmrpp'' application utilizes software from the Hyrax data server to produce the base DMR document which is used to construct the ''dmr++'' file.

The Hyrax server has a long list of configuration options, several of which can substantially alter the the structural and semantic representation of the dataset as seen in the dmr++ files generated using these options.

===Command line options===

The command line switches provide a way to control the output of the tool. In addition to common options like verbose output or testing modes, the tool provides options to build extra (aka 'sidecar') data files that hold information needed for CF compliance if the original HDF5 data files lack that information (see the ''missing data'' section ). In addition, it is often desirable to build ''dmr++'' files before the source data files are uploaded to a cloud store like S3. In this case, the URL to the data may not be known when the ''dmr++'' is built. We support this by using placeholder/template strings in the ''dmr++'' and which can then be replaced with the URL at runtime, when the ''dmr++'' file is evaluated. See the '-u' and '-p' options below.

====Inputs====

; -b
: The fully qualified path to the top level data directory. Data files read by ''get_dmrpp'' must ride in the directory tree rooted at this location and their names expressed as a path relative to this location. The value may not be set to <code>/</code> or <code>/etc</code> The default value is /tmp if a value is not provided. All the data files to be processed must be in this directory or one of its subdirectories. If ''get_dmrpp'' is being executed from same directory as the data then <code>-b `pwd`</code> works as well.
;-u
: This option is used to specify the location of the binary data object. It’s value must be an http, https, or file (file://) URL. This URL will be injected into the dmr++ when it is constructed. If option -u is not used; then the template string <code>OPeNDAP_DMRpp_DATA_ACCESS_URL</code> will be used and the dmr++ will substitute a value at runtime.
;-c
:The path to an alternate bes configuration file to use.
;-s
:The path to an optional addendum configuration file which will be appended to the default BES configuration. Much like the site.conf file works for the full server deployment it will be loaded last and the settings there-in will have an override effect on the default configuration.

====Output====

; -o
: The name of the file to create.

====Verbose Output Modes====

; -h
: Show help/usage page.
; -v:
: verbose mode, prints the intermediate DMR.
; -V
: Very verbose mode, prints the DMR, the command and the configuration file used to build the DMR, and does not remove temporary files.
; -D
: Just print the DMR that will be used to build the DMR++

====Tests====

; -T
: Run ALL hyrax tests on the resulting dmr++ file and compare the responses the ones generated by the source hdf5 file.
; -I
: Run hyrax inventory tests on the resulting dmr++ file and compare the responses the ones generated by the source hdf5 file.
; -F
: Run hyrax value probe tests on the resulting dmr++ file and compare the responses the ones generated by the source hdf5 file.

====Missing Data Creation====

; -M
: Build a 'sidecar' file that holds missing information needed for CF compliance (e.g., Latitude, Longitude and Time coordinate data).
; -p
: Provide the URL for the Missing data sidecar file. If this is not given (but -M is), then a template value is used in the dmr++ file and a real URL is substituted at runtime.
; -r
: The path to the file that contains missing variable information for sets of input data files that share common missing variables. The file will be created if it doesn't exist and the result may be used in subsequent invocations of get_dmrpp (using -r) to identify the missing variable file.

===''hdf5_handler'' Configuration===

Because ''get_dmrpp'' uses the ''hdf5_handler'' software to build the ''dmr++'' the software must inject the ''hdf5_handler'''s configuration.

The default configuration is large, but any valued may be altered at runtime.

Here are some of the commonly manipulated configuration parameters with their default values:
H5.EnableCF=true
H5.EnableDMR64bitInt=true
H5.DefaultHandleDimension=true
H5.KeepVarLeadingUnderscore=false
H5.EnableCheckNameClashing=true
H5.EnableAddPathAttrs=true
H5.EnableDropLongString=true
H5.DisableStructMetaAttr=true
H5.EnableFillValueCheck=true
H5.CheckIgnoreObj=false

====''Note to DAACs with existing Hyrax deployments.''====

If your group is already serving data with Hyrax and the data representations that are generated by your Hyrax server are satisfactory, then a careful inspection of the localized configuration, typically held in /etc/bes/site.conf, will help you determine what configuration state you may need to inject into ''get_dmrpp''.

===The ''H5.EnableCF'' option===

Of particular importance is the ''H5.EnableCF'' option, which instructs the ''get_dmrpp'' tool to produce [https://cfconventions.org/ Climate Forecast convention (CF)] compatible output based on metadata found in the granule file being processed.

Changing the value of ''H5.EnableCF'' to '''''false''''' will have (at least) two significant effects.

It will:

* Stop ''get_dmrpp'' from attempting to make the results CF compliant.
* Make visible the Group hierarchies (if any) in the underlying data granule as these would have been suppressed otherwise, as CF does not yet support Groups.

By default ''get_dmrpp'' the ''H5.EnableCF'' option is set to true:
H5.EnableCF = true

===Missing data, the CF conventions and ''hdf5''===

Many of the ''hdf5'' files produced by NASA and others do not contain the domain coordinate data (such as latitude, longitude, time, etc.) as a collection of explicit values. Instead information contained in the dataset metadata can used to reproduce these values.

In order for a dataset to be Climate Forecast (CF) compatible it must contain these domain coordinate data values.

The Hyrax ''hdf5_handler'' software, utilized by the ''get_dmrpp'' application, can create this data from the dataset metadata. The ''get_dmrpp'' application places these generated data in a “sidecar” file for deployment with the source ''hdf5/netcdf''-4 file.

==Hyrax - Serving data using dmr++ files==

There are three fundamental deployment scenarios for using dmr++ files to serve data with the Hyrax data server.

This can be simple categorized as follows:
The dmr++ file(s) are XML files that contain a root <tt>dap4:Dataset</tt> element with a <tt>dmrpp:href</tt> attribute whose value is one of:
# An http(s):// URL referencing to the underlying granule files via http.
# A file:// URL that references the granule file on the local filesystem in a location that is inside the BES' data root tree.
# The template string <tt>OPeNDAP_DMRpp_DATA_ACCESS_URL</tt>

Each will discussed in turn below.

'''Note''': By default Hyrax will automatically associate files whose name ends with ".dmrpp" with the dmr++ handler.

===Using dmr++ with http(s) URLs===

If the dmr++ files that you wish to serve contain <tt>dmrpp:href</tt> attributes whose values are http(s) URLs then there are 2+1 steps to serve the data:
# Place the dmr++ files on the local disk inside of the directory tree identified by the <tt>BES.Catalog.catalog.RootDirectory</tt> in the BES configuration
# Ensure that the Hyrax AllowedHosts list is configured to allow Hyrax to access those target URLs. This can be accomplished by adding new regex entires to the AllowedHosts list in /etc/bes/site.conf, creating that file as need be.
# If the data URLs require authentication to access then you'll need to configure Hyrax for that too.

===Using dmr++ with file URLs===

Using dmr++ files with locally held files can be useful for verifying that dmr++ functionality is working without relying on network access that may have data rate limits, authenticated access configuration, or security access constraints. Additionally, in many cases the dmr++ access to the locally held data may be significantly faster than through the native netcdf-4/hdf5 data handlers.

In order to use dmr++ files that contain file:// URLs:
# Place the dmr++ files on the local disk inside of the directory tree identified by the <tt>BES.Catalog.catalog.RootDirectory</tt> in the BES configuration
# Ensure the the dmr++ files contain only file:// URLs that refer to data granule files inside of the directory tree identified by the <tt>BES.Catalog.catalog.RootDirectory</tt> in the BES configuration.

Note: For Hyrax, a correctly formatted file URL must start with the protocol <tt>file://</tt> followed by the full qualified path to the data granule, for example:
/usr/share/hyrax/ghrsst/some_granule.h5
so the the completed URL will have three slashes after the first colon:
file:///usr/share/hyrax/ghrsst/some_granule.h5

===Using dmr++ with the template string.===

Another way to server dmr++ files with Hyrax is to build the dmr++ files '''without''' valid URLs but with a template string that is replaced at runtime. If no target URL is supplied to get_drmpp at the time that the dmr++ is generated the template string: <tt> OPeNDAP_DMRpp_DATA_ACCESS_URL</tt> will added to the file in place of the URL. The at runtime it can be replaced withe the correct value.

Currently the only implementation of this is Hyrax's NGAP service which, when deployed in the NASA NGAP cloud, will accept "restified path" URLs that are defined as
having a URL path component with two mandatory and one optional parameters:
MANDATORY: " /collections/UMM-C:{concept-id} "
OPTIONAL: "/UMM-C:{ShortName} '.' UMM-C:{Version} "
MANDATORY: "/granules/UMM-G:{GranuleUR}"
'''Example:''' <tt>https://opendap.earthdata.nasa.gov/collections/C1443727145-LAADS/MOD08_D3.v6.1/granules/MOD08_D3.A2020308.061.2020309092644.hdf.nc</tt>

When encountering this type of URL Hyrax will decompose it and use the content to formulate a query to the NASA CMR in order to retrieve the data access URL for the granule and for the dmr++ file. It then retrieves the dmr++ file and injects the data URL so that data access can proceed as described above.

[https://wiki.earthdata.nasa.gov/display/DUTRAIN/Feature+analysis%3A+Restified+URL+for+OPENDAP+Data+Access More on the Restified Path can be found here],

== Recipe: Building and testing dmr++ files using Hyrax docker containers ==
Prerequisites:
* Docker daemon running on a system that also supports a shell (the examples use <tt>bash</tt> in this section)
=== Recipe: Building dmr++ files using a Hyrax docker container.===
# Acquire representative granule files for the collection your wish to import. Put them on the system that is running the Docker daemon. For this recipe we will assume that these files have been placed in the directory:
#: <tt>/tmp/dmrpp</tt>
# Get the most up to date Hyrax docker image:
#: <tt>docker pull opendap/hyrax:snapshot</tt>
# Start docker container, mounting your data directory on to the docker image at <tt>/usr/share/hyrax</tt>:
#: <tt>docker run -d -h hyrax -p 8080:8080 --volume /tmp/dmrpp:/usr/share/hyrax --name=hyrax opendap/hyrax:snapshot</tt>
# Login to the docker container:
#: <tt>docker exec -it hyrax /bin/bash</tt>
# Change working dir to data dir:
#: <tt>cd /usr/share/hyrax</tt>
# Get a first view of your data using the default <tt>get_dmrpp</tt>. This script assumes your ingestable data files end with <tt>.h5</tt>, and the resulting dmr++ files will contain <tt>file://</tt> URLs so that they may be easily tested with the docker instance.
#: <tt>#!/bin/bash</tt>
#: <tt>for file in *.h5</tt>
#: <tt>do</tt>
#:: <tt>dmrpp_file="${file}.dmrpp"</tt>
#:: <tt>get_dmrpp -b `pwd` -o "${dmrpp_file}" -u "file://"`pwd`"/${file}" "${file}"</tt>
#: <tt>done</tt>
# Use the running Hyrax server to look at the dmrpp files you just created by pointing your browser at:
#: <tt>http://localhost:8080/opendap/</tt>

=== Testing and qualifying dmr++ files ===
In the previous section/step we created some initial dmr++ files using the default configuration. It is crucial to make sure that they provide the representation of the data that you and your users are expecting, and that they will work correctly with the Hyrax server. (See the following sections for details). If the generated dmr++ files do not match expectations then the default configuration of the <tt>get_dmrpp</tt> may need to be amended using the <tt>-s</tt> parameter.
If the data are currently being served by your DAAC's on-prem team this is where understanding exactly what the localizations made to the configurations of the on-prem Hyrax instances deployed for the collection is important. These localization will probably need to be injected into <tt>get_drmpp</tt> in order to produce the correct data representation in the dmr++ files.

=== Flattening Groups ===
By default <tt>get_dmrpp</tt> will preserve and show group hierarchies. If this is not desired, say for CF-1.0 compatibility, then you can change this by creating a small amendment to <tt>get_dmrpp</tt>'s default configuration.
First create the amending configuration file:
echo "H5.EnableCF=true" > site.conf
Then, change the invocation of <tt>get_dmrpp</tt> in the above example by adding the <tt>-s</tt> switch:
get_dmrpp -s site.conf -b `pwd` -o "${dmrpp_file}" -u "file://"`pwd`"/${file}" "${file}"
And re-run the dmr++ production as shown above.

==== Unwanted Side Effects ====
A side effect of setting <tt>H5.EnableCF=true</tt> is that some data types that are not represented in the DAP2 (and possibly early CF conventions) will be promoted to a different type, as is the case with signed bytes which are promoted to signed 16bit integers. Or, as in the case of 64bit integers, they may be elided from the dataset.

The promotion of variables in the creation of the dmr++ will create runtime problems and must, at least for now, be studiously avoided. Much of the legacy data being ripped to dmr++ may benefit from the default configuration of <tt>H5.EnableCF=true</tt> and the variable promotion behavior will not be an issue. However, for many of the more recently developed datasets this promotion must be carefully avoided.

How does one tell this is an issue for a particular dataset/collection?
: ''By using either the <tt>ncdump</tt> or the <tt>h5dump</tt> command one can see the native datatypes on the file.''

Example (''semantic metadata elided for readability/compactness''):
ncdump -h 20020601090000-JPL-L4_GHRSST-SSTfnd-MUR-GLOB-v02.0-fv04.1.nc
netcdf \20020601090000-JPL-L4_GHRSST-SSTfnd-MUR-GLOB-v02.0-fv04.1 {
dimensions:
time = 1 ;
lat = 17999 ;
lon = 36000 ;
variables:
int time(time) ;
float lat(lat) ;
float lon(lon) ;
short analysed_sst(time, lat, lon) ;
short analysis_error(time, lat, lon) ;
byte mask(time, lat, lon) ;
byte sea_ice_fraction(time, lat, lon) ;
We can see two variables, <tt>mask</tt> and <tt>sea_ice_fraction</tt>, whose type is signed byte. The default setting of <tt>H5.EnableCF=true</tt> will cause these two variables to be promoted in the resulting dmr++ to <tt>short</tt> (aka <tt>Int16</tt>) and that spells trouble for the dmr++ service. In "normal/legacy" server operations the promotion and subsequent type casting of the data values are handled in a single operation by either the ''netcdf_handler'' or the ''hdf5_handler''. But in the dmr++ operation these activities happen at separate times, and the downstream code knows not of the promotion which will cause data corruption issues in the servers responses. The fix is easy, rerun <tt>get_dmrpp</tt> with the <tt>-s</tt> switch and inject <tt>H5.EnableCF=false</tt> to quell this behavior.

=== DAP representations ===
We have test and assurance procedures for DAP4 and DAP2 protocols below. Both are important. For legacy datasets the DAP2 request API is widely used by an existing client base and should continue to be supported. Since DAP4 subsumes DAP2 (but with somewhat different API semantics) It should be checked for legacy datasets as well. For more modern datasets that content DAP4 types such as Int64 that are not part of the DAP2 specification or implementations we will need to relying eliding the instances of unmapped types, or return an error when this is encountered.
# Test Constants:
GRANULE_FILE="some_name.h5"
# Granule URL
gf_url="http://localhost:8080/opendap/${GRANULE_FILE}"

==== Inspect the dmr++ files====
# Do the dmr++ files have the expected dmrpp:href URL(s)?
#: <tt>head -2 ${GRANULE_FILE}.dmrpp</tt>

==== Check DAP4 DMR Response ====
Inspect <tt>${gf_url}.dmrpp.dmr</tt>
# Get the document, save as <tt>foo.dmr</tt>:
#: <tt>curl -L -o foo.dmr "${gf_url}.dmr"</tt>
# Is each variable's data type correct and as expected?
# Are the associated dimensions correct?

==== DAP4 Check binary data response ====

For a particular granule GRANULE_FILE and a particular variable VARIABLE_NAME (Where [https://docs.opendap.org/index.php?title=DAP4:_Specification_Volume_1#Fully_Qualified_Names VARIABLE_NAME is a full qualified DAP4 name.]):
:<tt> curl -L -o dap4_subset_file "${gf_url}.dap?dap4.ce=VARIABLE_NAME"</tt>
:<tt> curl -L -o dap4_subset_dmrpp "${gf_url}.dmrpp.dap?dap4.ce=VARIABLE_NAME"</tt>
:<tt> cmp dap4_subset_file dap4_subset_dmrpp</tt>

==== DAP4 UI test ====
* View and exercise the DAP4 Data Request Form <tt>{gf_url}.dmr.html</tt>

==== DAP2 Check DDS Response ====

# Inspect <tt>${gf_url}.dds</tt>
## Is each variable's data type correct and as expected?
## Are the associated dimensions correct?
# Compare DMR++ DDS with granule file DDS.
#:For a particular granule GRANULE_FILE and a particular variable VARIABLE_NAME (Where [https://cdn.earthdata.nasa.gov/conduit/upload/512/ESE-RFC-004v1.1.pdf VARIABLE_NAME is a DAP2 name.]):
#::<tt> curl -L -o dap2_dds_file "${gf_url}.dds"</tt>
#::<tt> curl -L -o dap2_dds_dmrpp "${gf_url}.dds"</tt>
#::<tt> cmp dap2_dds_file dap2_dds_dmrpp </tt>

==== DAP2 Check binary data response ====

For a particular granule GRANULE_FILE and a particular variable VARIABLE_NAME (Where [https://cdn.earthdata.nasa.gov/conduit/upload/512/ESE-RFC-004v1.1.pdf VARIABLE_NAME is a DAP2 name.]):
:<tt> curl -L -o dap2_subset_file "${gf_url}.dods?VARIABLE_NAME"</tt>
:<tt> curl -L -o dap2_subset_dmrpp "${gf_url}.dmrpp.dods?VARIABLE_NAME"</tt>
:<tt> cmp dap2_subset_file dap2_subset_dmrpp</tt>

'''Note''': One might consider doing this with two or more variables.

==== DAP2 UI Test ====
* View and exercise the DAP2 Data Request Form located here: <tt>{gf_url}.html</tt>
* Try it in Panoply!
** Open Panoply.
** From the '''File''' menu select '''Open Remote Dataset...'''
** Paste the <tt>{gf_url}.html</tt> into the resulting dialog box.
---

2009-04-29T17:30:15Z

Dan:

Changes to the DAP, Hyrax, or other proposals stemming from the WCS Gateway Project.

Much of the WCS gateway work is focused on:

# Providing DAP2 protocol access to WCS served data.

=== DAP2 access to data served by WCS ===

* [[WCS_Data Access|Data access via DAP2 requests]]
* [[WCS Site Analysis |WCS Server Analysis]]
==== DAP2.1 Namespace ====
There have been some proposed changes to the DAP2 namespace (for the DDX) to hep accommodate this work.
* [[WCS DAP Attributes|DAP Attributes]]

----