DAP4: Async Service for AWS Glacier

From OPeNDAP Documentation
Jump to: navigation, search

1 Background

As part of the DAP4 specification we have defined a mechanism through which clients and servers can negotiate access to data resources that are "near-line". These data resources may be held in a tape drive system or some other system that takes an intermediate amount of time (a few hours to a few days) to extract a stored resource. The Amazon Web Services Glacier system is an excellent example of such a near-line system and we started developing a prototype service to provide DAP4 client access to Glacier holdings.

2 Overview

Here are some Glacier axioms (things which we hold to be true, even if they make life hard):

  1. The Glacier service is organized into "vaults" (a named container similar to the AWS S3 bucket) into which things (data resources, or in Glacier speak "archives" which are similar to the AWS S3 "object"s) are put
  2. The Glacier service typically takes about 4 hours to respond to a request for an "archive".
  3. The Glacier service builds an "inventory" of each vault's contents every 24 hours.
  4. Requests for the current vault inventory also take 4 hours to complete.
  5. When an archive is uploaded to Glacier the name (aka archive-id) under which it is stored is determined by Glacier and returned when the upload is finished. So if you ever want to get that archive back you either have to remember the archive-id locally or be prepared to find it in the not so easy to get inventory.
  6. When you upload an archive you can provide a description of it. The AWS documentation suggests that you use your own resourceId (for example a local path/file name string) so that if you do need to rebuild your local representation of the archive you can do so from a Glacier inventory (which contains both the archive-ids and the descriptions).
  7. Glacier will happily allow you to upload the same resource over and over again. Each time you upload the resource you will get a new archive-id for it. The previously uploaded copies will remain in the vault associated with their previously assigned archive-ids unless you actively delete them.
  8. Archives can be deleted.
  9. Vaults can be deleted.
  10. Uploading archives, deleting archives, and deleting vault operations do not require asynchronous behavior on the part of the client. In other words these operations happen in "real time".
  11. Retrieving archives and vault inventories require asynchronous behavior capabilities on the part of the client

(And the Real Numbers are defined with only 16 axioms)

3 Design Issues

In order to utilize Glacier as part of a DAP4 service several things need to be in place: A catalog system, some kind of database that associates glacier holdings with resources identified in the catalog, and a system for holding retrieved glacier archives/resources for use by the service.

3.1 Catalog

catalog
A directed graph that embodies a hierarchical representation of a collection of data resources.

Some abstract representation of the Glacier holdings will be required. While catalogs are not part of the DAP4 spec we know that in general DAP servers return catalogs in the form of THREDDS catalogs or as HTML pages that represent catalog nodes and that contain links to data resources and other catalog nodes. While a Glacier vault might simply be represented by a list of archive-ids this is not a representation that users of data resources would find very useful, especially when coupled with the delay in discovering anything about any single resource.

The catalog should be close to the server and quickly accessible.

3.2 Archive Records

ArchiveId
The string which is returned by Glacier to identify an uploaded archive and through which said archive may be retrieved from Glacier. These appear to be GUIDs. Example: XIfGXl0qWkfDVbKL3Ibb6XQ8OBC44YwOBGl9ds-yA18HN8zDOBi1A2ZugWprFHgwL-8LRQLSfz8BqoJLaMdaflPgM-J-_D0gHQwS_CDHXp5R0vHUADuZKd49oP4eir2qw508EJ57kQ
ResourceId
A string that is used by a DAP server to identify a particular (data) resource. Typically a path/file name, but in practice this would appear as a local path fragment in a DAP URL after the domain-port-service part of the URL. For example in the url http://test.opendap.org:8080/opendap/hyrax/data/nc/fnoc1.nc the resourceId would be data/nc/fnoc1.nc
Archive Record
A record (as in a file or database entry) that contains at minimum a resource's resourceId and archiveId

If all that was required was to associate the resourceId and archiveId it might seem reasonable to just include this small set of information in the persisted catalog for the Glacier vault. However because retrieving anything from Glacier costs time (and money) it is probably important to capture some kind of information about the each archived resource locally and in such a way that users can inspect resource metadata without requesting that the archived resource be retrieved just to find out that the resource is of little or no interest.

Suggestion: At minimum store the dap4:DMR (or dap2:DDX) document in the local Archive Record. Consider also storing the names and values of any MAP arrays found in the dataset. Maybe use the XML data response to generate this content as part of the uploading to Glacier process.

Example: Glacier Archive Record

3.3 Resource Cache

Resources that have been retrieved from Glacier must be put somewhere so that the DAP (and possibly other) services can utilize them. I call this a cache because I think it's implied that the volume of things in a Glacier vault may be much larger than any reasonable machine running this service may be able to hold. Therefore:

  1. The retrieved Glacier resources must be cached.
  2. Cached resources should be retrievable by utilizing their resourceId.
  3. The cache must have a size limit.
  4. There must be software that manages the cache contents by applying a metric to determine which resources should be purged from the cache and when.

4 Implementation (Partial + Ideas)

All access to Glacier must be authenticated and this argues for using one of Amazon's APIs instead of 'direct' access with HTTP GET, PUT, etc. Amazon provides Java and Python APIs (as well as some others) but notably not C/C++.

I have written about half of a prototype for this using the OLFS and the local file system. I create a directory for each vault and underneath this are 3 subdirectories: cache, index, archive.

4.1 cache directory

This is where I plan to store the resources retrieved from glacier. The intention is to store the retrieved resource content in a file whose name is the file system compatibility encoded resourceId. This way there is only a single directory to look in and when a request for a resource is being handled by the server the server only has to encoded the resourceId and check here to see if it's already got the thing.

An attractive alternative to this would be to use S3 bucket as the cache:

  • No encoding the resourceId's
  • Really big :)
  • The BES already knows how to read things from S3 via HTTP and the gateway.
  • The OLFS could handle managing the Glacier and S3 interactions and only forward DAP requests to the BES once the resource was available in S3.

4.2 index directory

(Note: Index and catalog may be synonymous here.)

The index directory is where I am keeping the index/catalog files. I am using as the index scheme developed by Jeff Orgata at NOAA NODC. I think it's probably important for for server performance for the index to be kept 'close by' - as in a locally mounted filesystem. (In an AWS instance I could see using an EBS volume). The index files are saved to the index directory using their file system encoded resourceId as the filename.

Thoughts:

  • Keeping all of the index files in a single directory with no other content allows the server to ingest the entire index by simply grabbing all the files in the directory. This can useful if the entire index is to held in a memory cache which may or may not unreasonable depending on the size on the index and the server resources that may be available. Certainly memory caching the index would increase traversal performance.
  • If other files are intermingled in the directory with the index files then a mass ingest is not possible and the top level index must be loaded and the graph traversed in order to load the index. If the index has multiple trees/roots then there would necessarily need to be multiple starting points for the traversal.
  • Saving the index files using their encoded resourceIds allows the server to lazy evaluate the catalog, simply loading each index as requested. The ingested index can be held in some kind of managed memory cache or discarded if the index traversal performance is not an issue.

4.3 archive directory

The archive directory is where I store the the archive records as described above. As with the index and cache directories, these records are stored in a file whose name is the file system compatibility encoded resourceId. These files represent things that could/may/will end up in the cache and the name they are stored under is the same name as the downloaded resource in the cache directory. That would be intentional: This way the server can use the resourceId to quickly determine if a resource exists (by attempting to retrieve an archive record), and then if it's already cached (by locating the cache file).

4.4 Thoughts

I made a choice to separate the cache, index, and archive contents. I could used an alternate scheme in which all these files get placed in the same directory on the file system - the resourceId's for the index and the data resources will be, be definition, different. The archive and cache files could be combined into a single directory by using a naming extension applied to the resourceId (ex: resourceId.cache & resourceId.arcrec) . I separated them into their own directories to help my own human eyes and brain, I think it's worth noting because by combining all of these things into a single directory you would be using the file systems as a convenient, heterogeneous, no-sql database. In fact one might see some utility in replacing the use of the filesystem here with a real no-sql database - but that's a story for another day.