Articles from Operations

  • Registry: A Janitor Speaks Out

    2023-04-25 Markus Demleitner

    I sometimes claim the reason I like working on the VO Registry is that I am a librarian at heart. Perhaps there is some truth to that, in that ugly metadata does make me unhappy – but beyond that, it also makes the Virtual Observatory look or even work a good deal worse than it should.

    Given that, in this post I'm afraid I will sound more like a grumpy janitor than a wise librarian, but let me still attempt to contribute to better metadata by pointing out a few things to watch out for when writing a resource record. People consuming resource records (i.e., VO-using astronomers) are welcome here, too: when you encounter antipatterns mentioned here, a polite complaint to the service publisher is entirely a good thing.

    Note that I am using real metadata found in the registry – in case you recognise some of own records, do not feel reprimanded individually. Most of the problems I discuss here are really common at this point, and thus if I picked your metadata, that was mere bad luck. I actually picked some of my own occasionally (but duly fixed the problem then).

    Missing Coverage

    Since VODataService 1.2, you can say what part of the sky, spectrum, and time your resource covers. That is incredibly useful metadata in practice. Spatial coverage, for instance, is used in Aladin like this:

    Screenshot: Resource names in white, orange and green, and a part of the sky (h and χ Persei) next to them

    Green means: these services could have data for the patch of sky shown, orange means don't bother with these, and white means: No idea because the resource does not declare its coverage.

    Similarly, it would be great if researchers or clients could reliably say:

    SELECT * FROM rr.resource JOIN rr.stc_spectral WHERE
  1=ivo_interval_overlaps(spectral_start, spectral_end,
      ivo_specconv(658, 'nm', 'J'), ivo_specconv(654, 'nm', 'J'))

    to find resources having data covering the Hα line on the spectral axis. Currently, that's just 2064 resources, and given that Hα sits smack in the middle of the optical window that's an indication that far too few resources say where they are.

    So – add STC coverage to your data today. It's not hard with pymoc or pgsphere and chapter 3.2 of VODataService. DaCHS operators will probably get by just studying the corresponding section of the tutorial.

    Broken Author Names

    On the ADS, last time I had information on that, about 90% of the queries were by author. In the VO registry, by my unscientific estimate, less than 5% of queries constrain authors. Sure, people look for literature and data in different ways and for different purposes, but an important reason for the difference still is that we don't do a good job giving creator/name (which contains the equvialent of the author name).

    The ideal format is to have last name first, then a comma, and then abbreviated initials or full first names, as in von der Heide, J.. Many names in the VO are almost in this format do not have a comma; but the comma makes parsing these names a lot simpler, so please put it in. Of all the forms to write names in, that's most easily constrained without guessing how many first names are where. Remember, there are people out their with names like „Kirsten-Claude Selim de Vaucouleurs-van der Heide Lobos“ (or, for that matter, Uthamadhanapuram Venkatasubbaiyer Swaminatha Iyer), and a computer cannot efficiently decide where the last name starts in first name first order (and conversely, without the comma in last name first order, it has a hard time figuring out where the last name stops). Also, last name first almost always gives a more useful natural sort order.

    Realistically, people will have to live with J. von der Heide, too, so author searches in the VO will have to look like LIKE '%von der Heide%' for some years to come, but let's at least try to improve. And whatever you do, don't do any of (in approximate order of severity):

    • Dump in half an acknowledgement, e.g., under a cooperative agreement with the NSF on behalf of the Gemini partnership: the National Science Foundation (United States), or, about as bad: provided by S. Snowden from data by Dickey and Lockman – that's useless for author searches but invites lots of false positives
    • Dump more than one name into one creator/name element, e.g., Zhuang Z.,Kirby E.N.,Leethochawalit N.,de los Reyes M.A.C. or Voges, W.; Aschenbach, B.; Boller, Th.; (and ~200 more characters) – that's really hard to search and essentially impossible to use for, e.g., author datagraphies
    • Include affiliations (the VO can't properly deal with those yet), e.g., Zub M. (The PLANET Collaboration) or a combination of this and the previous: Zhu W. (The Spitzer team) Dominik M.
    • Forget citation debris, e.g., et al. MNRAS (in press), or, shockingly common: and Scheck M.; of course, entire citations (WALKER I. Astron. J. 106) are inappropriate, too – all of this will prevent the use of meaningful name constraints
    • Give a bibcode: 2014ApJ...787...78M – this likely belongs into content/source
    • Have empty author name elements (as, at this moment, 13 records)
    • Cheat with effectively empty author names: <NOT GIVEN>, or "We forgot to give credit, please complain"
    • Go all uppercase, e.g., ZINNECKER H. – standards-compliant ADQL string comparisons are case-sensitive, and case-folding would require special indexes. Perhaps case-insensitive author matches should be made easier in that van der Waals is probably the same person as Van der Waals, but for now that's not how it works right now. And I don't think that will change any time soon, because if I have learned one thing in my life it is that case insensitivity is almost always evil
    • Have just a first name: walter or W.I. or W-J
    • Combine author lists from different contributing papers: Wright et al.; Griffith, Wright, Burke, Ekers; Griffith, Wright – if you really need to do something like this, merge the two author lists – and then of course use one name per creator element

    In principle, these considerations would apply to contributors, contacts and perhaps publishers, too, but since I don't think people should use these in discovery queries, their format does not matter too much: If they're human-readable, that's enough.

    Fragile Contact Info

    Quite regularly I need to ask people to fix something in their publishing registries, and then it's really useful to have reliable contact information. That's also nice for VO users; pyVO, for instance, has the get_contact method on registry records, and in WIRR, you can pop up contact info on all records:

    a screenshot showing a match in a registry query. A subwindow is popped up that shows a mail address and a telephone number of a “GAVO Data Center Team“.

    For that to work, personal addresses in the contact information are really dangerous – it is my experience that these break significantly more often than institutional addresses. So, please avoid things like (I'm making all of these up because there may still be folks around harvesting mail addresses to send spam):

    • a.b.miller-parachtnix@gmail.com (well: avoid using gmail.com unconditionally)
    • friederike.student@ari.uni-heidelberg.de

    Rather, create an alias that you can hand on and that perhaps is even a bit speaking. This could be:

    • vo-help@great-telescope.org
    • gavo@ari.uni-heidelberg.de
    • uni-hd-vo@posteo.de (in case your own institution absolutely loathes the idea of addresses not bound to persons)

    Non-machine-readable Subjects

    VOResource 1.1 said that subjects are to be taken “from the UAT” (that's the Unified Astronomy Thesaurus), but failed to say what exactly that means. Since last July, this is properly defined: Use fragment identifiers into http://www.ivoa.net/rdf/uat, that is, something like abell-clusters.

    Having all subject keywords in a predictable format, with useful metadata, and part of a proper hierarchy enables all kinds of cool stuff, and hence it would be great if we could stomp out the following sorts of mispractice in the VO:

    • Multiple things in one subject element: ATLAS DR1, SIAP, Images – have one term per subject element
    • Undefined NULL values: NOT PROVIDED, ??? – if you really cannot find a pertinent term, use astronomical-research (or one of the other top-level terms). If nowhere else, that at least helps when your record moves to interdisciplinary search engines
    • Random free text: optical lines equivalent width catalog – that's multiple terms rolled into one, and the machine will not know what it means
    • Project or instrument names: 6dF Data Release 3 Spectra, COROT N2 – there's the instrument metadata for some uses of that. For the rest, see above on having projects in creator/name.
    • Protocol names: TAP – that's what capabilities are for
    • Service titles: CADC image/cube HiPS service – that's what the title element is for
    • Non-UAT keyword schemes: Galaxy:general – let's not force VO components to learn about multiple keyword systems. If you are missing something from the UAT, tell them about it

    Unfulfilling Resource Descriptions

    Descriptions of VO resources serve a dual purpose: The should give researches a quick idea of what to expect and not expect of a resource, and they should mention all the important buzzwords for the benefit of full-text searches. Hence, if you only have two words as in:

    Survey (LoLSS).

    or have something like a title:

    Convolution of normalized synthetic stellar spectra.

    or use somewhat uncommon abbreviations and technical details that probably will not help much during data discovery:

    USET Group form

    (what group? Does „form“ really mean „web browser-facing“? If so, that's again better expressed through the capabilities), you should work a bit on your description.

    It is usually helpful to start the description with „this service is…“ or something similar. While it's marginally ok to mention terms and conditions like:

    When referencing results from this online catalog, please cite &lt;a href="https://iopscience.iop.org/article/10.384…

    further down in the description (the proper place for this kind of thing is the rights element, though), don't discuss stuff like this before you have told people what is in the “online catalog” in the first place. Also: registry records are like e-mail in that you shouldn't use HTML anywhere in registry metadata. If you have to include URLs in text for human consumption, just put them in as text.

    Talking about markup: You cannot rely on any of that in descriptions. Even basic ASCII art (or, well, tables) will always come out bad:

    Only the data from the first catalog that was matched is reported here according to the following priority list. This means for example, if a star was matched with Hipparcos, that information was used while possible other catalog data are not listed here. -------------------------------------------------------- # stars flg catalog -------------------------------------------------------- 53500 0 no catalog match 55549 1 Hipparcos 254 2 Yale Parallax Catalog 1041 3 Finch and Zacharias 2016 (UPM NNNN-NNNN) 1431 4 MEarth parallaxes 402 5 SIMBAD Database (w/parallax) -------------------------------------------------------- 112177 total number stars in catalog -------------------------------------------------------- Not all parallaxes from the...

    (of course, that in this case the newlines and longer sequences of blanks have been normalised to single blanks already in the original resource record makes it particularly certain that the table will come out wrong).

    And where in titles abbreviations are usually a good thing, in particular when you can expect your target audience too look for the abbreviation rather spelled-out names in discovery queries, in descriptions you have space, and hence you normally should explain MCQA as „Monte Carlo Quality Assessment“ in something like the following:

    Herschel sources in Planck fields measured at 350 µm MCQA

    Remember: The people who read your descriptions may come from the future (as in: 25 years from now) or at least may be unfamilar with your project's jargon.

    Lame Relationships

    There are an incredible 136958 relationships in the current VO that have related-to as their relationship type. This is deplorable because the relevant vocabulary, https://www.ivoa.net/rdf/voresource/relationship_type, marks it as deprecated, and that's for a good reason: Just stating “some relationship“ between two resources is rarely useful. Decide what the relationship is and then pick a proper term (or, if that does not exist, prepare a VEP).

    Missing Tablesets

    Tablesets are a VODataService feature giving metadata on the return table (or, in the case of the flexible TAP services, the queried tables). They are really useful if you look for services returning some sort of physics – and if you are running TAP services, they will one day let me shut down the GloTS service that replicates a good deal of registry functionality for no good reason at all.

    So, if you have a catalog service and your registry record ends somewhat like this:

      </capability>
</ri:Resource>

    it is almost certainly missing a tableset (which would normally go after the capabilities; you are probably also missing the sky coverage, though, because that would sit there, too).

    Writing basic tablesets is not hard. In fact, if you are running a TAP service, you have a working tableset on your service's tables endpoint. But even without VOSI tables, making a tableset from the VOTable you return is straightforward – with a few encouraging words, I could be talked to write a few lines of Python that do that.

    I will readily admit that writing good tablesets is more involved, but what is hard about it you should be doing anyway, because it also will improve the VOTables that you write, and hence the usability of your data all around. So, until the end of this post let me look at some common warts of the column metadata in today's VO.

    Deficient Column Descriptions

    Column descriptions like ?, ??, or even ??? are surprisingly common. Please don't do that. If you really have no idea what your upstream has put into a column, admit that, aplogise and try to make your upstream explain.

    And while RA somewhat works among astronomers, a word or two on the reference system (“IRCS”) and an informal provenance (“from PSF fits”) would certainly be much appreciated by your users and might even come handy in discovery.

    Or consider “Age” – this could immediately be improved by revealing just what has aged here and, again, some hint on how the age was estimated (e.g., “obtained from ivo://foo.bar/res” versus “by isochrone fitting”).

    But don't overdo it, either: Do not include entire footnotes in descriptions, because that will lead to many false positives in full text searches (not to mention slow down the Registry as a whole if this became common practice). DaCHS operators: you can have footnotes in your RD by using note meta items; cf. Typed Meta Elements in the DaCHS reference.

    Near the upper limit of what is appropriate in a column description is perhaps something like this:

    The 2.5 percentile of the Log total SFR PDF. This is derived by combining emission line measurements from within the fibre where possible and aperture corrections are done by fitting models ala Gallazzi et al (2005), Salim et al (2007) to the photometry outside the fibre. For those objects where the emission lines within the fibre do not provide an estimate of the SFR, model fits were made to the integrated photometry.

    – but at the same time it illustrates how you can provide a lot of information that helps casual users.

    The position angles I will turn to in a second give another nice example of why human-readable descriptions are so important: There is no reliable convention of the direction and the baseline of these, so stating something like „north over east“ in a description will avoid a lot of head-scratching.

    Column UCDs: Missing, Outdated, or Useless

    A very plausible discovery scenario involves UCDs: „give me resources with (some photometry | redshifts | kinematics | dynamics | positions on earth)“. Hence, make sure your columns' metadata has predictable and halfway correct UCDs.

    Sure, that's not always straightforward (note, by the way, that there is a reasonably simple process to suggest new UCDs), but there's no excuse for there being 117 columns called pa without any UCD, where pos.posAng will almost certainly fit all of them (though, who knows: 30 of these in addition don't even have a description).

    To make sure the UCDs you assign exist, run them through astropy at least once. Do not ignore complaints by astropy; it is actually preferable to have no UCD rather than “??” (which currently a whopping 30342 column sport, in addition to which we have 41 times “???“ and 70 times “????“[1]). Also, resist the temptation to freely invent things, such as the “mjd” UCD I'm seeing on 13 columns. In this particular case, by the way, I give you that saying “this column contains MJDs“ has been a pain in VOTables for a long time, but since version 1.4, TIMESYS lets you do that in a reasonable way.

    Oh, let me qualify the “freely invent“ in the last paragraph: It could be[2] that MJD has actually been part of the original UCDs you may still know from cone search (“POS_EQ_RA”); that people have not updated their metadata from these ancient days is also the reason I'm still seeing 13827 columns with an (invalid) UCD of “error“ in column metadata (and 84 with pos_eq_dec).

    Unrelatedly (though with an undisputable entertainment value): the longest UCD in the current VO is meta.code;phot.flux.density;arith.ratio;em.ir.15-30um;em.radio.750-1500mhz; unless I and astropy are missing something, it's even syntactically correct.

    Bad Units

    While I do not see many discovery scenarios that would make good use of units, do not forget to update your units to VOUnits when you touch up your tablesets. This will let software like astropy do the unit calculus for its users, which is a win overall. It cannot do that if you ignore VOUnits and write, say, ABmag/arcsec2 – the AB part you will have to communicate in the description for now, and exponentiation is ** in VOUnits.

    Recent versions of the stilts validators (votlint, taplint) will complain about bad units. And you can use stilts interactively to figure out whether you got it right:

    $ stilts calc 'vounitStatus("ABmag/arcsec2")'
  BAD_SYNTAX
$ stilts calc 'vounitStatus("mag/arcsec**2")'
  OK

    [In a previous version of this post, I have given a piece of astropy to do unit checking; it turns out that astropy by default is rather forgiving, and you want stilts on your box anyway; why not use it for unit validation? If your stilts says something about “bad expression“ with the command lines above, it's an indication that you should update it.]

    And with this somewhat non-registry topic: Go forth and polish your resource records. Or, as a consumer of such metadata, ask the publishers of bad resource metadata to fix it.

    [1]Remarkably, there are no ????? or even longer sequences of question marks, and even more remarkably, nobody has put in a lonely question mark. If someone versed in cognitive psychology has a plausible interpretation for that fact: can you share it with me?
    [2]Since the original UCDs predate my VO involvement and, for all I know, never were properly standardised, I frankly can't say.

  • Migrating Away From Wordpress

    2021-10-20 Markus Demleitner

    Since 2016, this blog was served through a Wordpress instance at the Astrophysical Institute Potsdam AIP – thanks again to our colleagues there for maintaining the platform over all these years.

    But since it now seems as if this is something that might last a long time (by Web standards), we have decided that we should leave PHP behind and look for something properly version controllable, and something that can simply live somewhere on a web server with essentially zero maintenance. Hence, we have moved the content to pelican – which has a clean Debian package, is written in Python, and does not need any active components of its own.

    As an extra bonus, the blog posts are now authored in ReStructuredText, which happens to be what DaCHS' documentation is written in, and what you can use to author metadata for DaCHS resources. If you want, you can now check out the source code for the articles (sorry, it's still subversion; one of these days I'll find something fancier than naked git but lighter than gitlab, and then I'll move GAVO's VCS to git).

    As expected, porting the theme (which I only did rather half-heartedly, so things are a bit less pretty now) and getting the figures right was what caused the bulk of the work. On the plus side, I have also greatly cleaned up categories and tags. Still, it's quite likely we messed something up. If you find anything broken here, please let us know: https://www.g-vo.org/pmwiki/About/Impressum lists the main ways through which you can reach us.

    With that: Subscribe to our Atom feed!

  • Taming the Postgres JIT

    2021-09-23 Markus Demleitner

    Mild warning: This is exclusively technobabble mainly addressing DaCHS deployers. If you're an astronomer (or yet something else), you're of course still welcome to enjoy it, but don't complain if you're bored.

    My development machine as been on Debian bullseye for a while, which means I've been running Postgres 13 for the past few months. Against Postgres 11, 13 is a lot more optimistic when doing Just-In-Time (JIT) compilation, and that's the beginning of this story.

    This JIT thing in plain language means that Postgres is writing small programmes to compute query results, then compiles them to machine code and executes that rather than running the query plan in some sort of interpreter. This at first sounds like a great idea that should speed up large queries quite a bit. But for one, query time is often bounded not so much by CPU but by I/O, and the sort of analysis that happens for JIT compilation is not free. Not at all.

    I noticed that when a query in the regression test suite I'm running before every commit to DaCHS started to occasionally fail. That test executes:

    SELECT TOP 1 obs_publisher_did
FROM ivoa.obscore
WHERE distance(s_ra, s_dec, 83.8,-5.4)<0.2

    and then asserts that the result is in within 10 seconds. The purpose of this particular regression test is to make sure all sizable tables in the obscore view have a usable spatial index on the production system. On the development system, there really aren't any tables in obscore that would be slow even when seqscanned.

    How on earth could this query be slow then?

    The natural reaction in such a situation to use EXPLAIN in psql. In this case, there is some non-trivial rewriting of the query going on between ADQL and postgres, which means you cannot just paste the ADQL to Postgres. To figure out the query that DaCHS actually executes, I picked the translated query from the VOTable returned from a successful request (look for the sql_query INFO; that's a DaCHS extension, so that trick won't work for other TAP servers), ran the psql gavo DaCHS operators are probably used to, and then typed:

    EXPLAIN SELECT obs_publisher_did
FROM ivoa.obscore
WHERE  q3c_join(83.8, - 5.4, s_ra, s_dec, 0.2) LIMIT 1;

    to it. The result was inconspicuous; a few seqscans here and there, but the total cost estimate was “0.00..7.12”, which in physical units works out to “basically nothing”, many orders of magnitude away from the 10 seconds I occasionally saw in the regression tests.

    Well, when a query plan doesn't match your expectations, the next thing to do is EXPLAIN ANALYZE. With that, Postgres executes the plan it has made and then compares its estimates to what the cost turned out to be; this, by the way, is also a good way to find out when you should raise the statistics target of one or more of your columns (see Element Column in the DaCHS reference for details).

    For me, the output looked something like this:

    Limit  (cost=10000000000.00..10156565675.53 rows=1 width=57) (actual time=6206.883..6206.899 rows=1 loops=1)
[...]
 Planning Time: 22.174 ms
 JIT:
   Functions: 130
   Options: Inlining true, Optimization true, Expressions true, Deforming true
   Timing: Generation 55.404 ms, Inlining 107.280 ms, Optimization 3479.626 ms, Emission 2601.411 ms, Total 6243.721 ms
 Execution Time: 6263.243 ms

    Ok, I'm lying a bit; there is another reason than just the analyze for why the cost estimate exploded from 7.12 to 10156565675.53. I'll confess in the appendix to this post.

    The main point, however, is: the execution time now is of the order that I'm expecting (the database is rather busy during a regression test, so those 6 seconds can easily become double that then). Interestingly, essentially all the execution time went into “Optimization” and “Emission”. Until yesterday, I'd never seen a thing like that in Postgres query plans.

    That is because here the JIT is at work, and that was at least a lot less likely in Postgres 11. Now, estimating 10 Gigapennies as execution cost up front, Postgres 13 thought some extra time for writing and compiling a little programme is well spent. Of course, that estimate is badly off, and the right thing to do is to fix the reason for the bad estimate. See the appendix for why I don't just yet.

    That my obscore view has 32 tables contributing to it, giving its definition a whopping 1280 lines, probably does not help. But in particular since the query plans in the presence of Q3C and pgsphere still are usually badly off, it might be wise to discourage Postgres a bit from using JIT compilation with DaCHS' workloads in your configuration if you're running TAP services (you should) and before you upgrade to Postgres 13. To do that, add a:

    jit_above_cost = 20000000000

    (or so; perhaps you can set your limit a good deal lower) to your postgresql.conf. On Debian boxes, that file is in /etc/postgresql/13/main/ (obviously, change the 13 if you have a different version). You need to restart postgres to make this take effect.

    While I was in that file, I thought I can share what other configuration I have in there, because it is likely you can speed up your data centre quite a bit by judicious tuning. The following settings aren't particularly well thought out, but I claim they are not unreasonable for a 64 GB machine that runs as a dedicated server; that last thing also causes the first configuration item, as for two-server operation, you have to set

    • listen_addresses = '*' – only then can you talk to postgres from another machine (disregarding hacks like ssh tunnels that may even work as last-resort options). Of course, this may mean your postgres port is visible to the internet, which means you ought to understand what pg_hba.conf is before configuring that. Other configuration I'm doing includes
    • max_connections = 200 – I actually ran out of connections once; DaCHS itself is now a bit more parsimonious with them, but if you have enough RAM, it still doesn't hurt to be generous here.
    • localtime = UTC – TIMESTAMPs suck, because it is hard to compute with them, are a pain when plotting, there are time zones, and they generally are a Babylonian mess (as evinced by base-60 numbers). But you can't always escape timestamps, and if you somehow manage to create them “with time zone”, telling the server to do UTC helps limit their damage radius.
    • shared_buffers = 15GB – the Postgres documentation says 25% of the RAM is a good default for shared_buffers, so that's roughly what I went for here. Note that the kernel usually limits how much shared memory processes are allowed to allocate, and you will have to adjust those limits for this to take effect. On Debian, the postgresql-common package installs a file /etc/sysctl.d/30-postgresql-shm.conf for easy adjusting of the limits.
    • temp_buffers = 100MB – that one gives buffers for temporary tables, and raising it helps TAP uploads (which use those, at least for now). Since our TAP uploads tend to be large as temporary tables go, it pays to set aside a couple of megabytes for them. Now that I look at this again and think about what people upload into my data centre: I think I could even raise that a bit more.
    • work_mem = 64MB – this one is for doing joins and the like (which includes cross-matches), and again these tend to be larger in Astronomy than in many other disciplines, where matching tens-of-millions against billions would count as Big Data. Hence, postgres' default of 4 MB is quite certainly going to be causing a lot of unnecessary disk activity. That said, DaCHS could be a bit smarter here and raise work_mem itself when running TAP jobs (or perhaps only TAP jobs that actually do joins). Note that a single query can use up many times work_mem, which means you shouldn't choose this too high, either. One thing I'd like to look into one day is the hash_mem_multiplier (cf. a bit down on Postgres docs on resource limits). If you do research in that direction with astronomy workloads, please let me know.
    • maintenance_work_mem = 2048MB – this is relevant to keep VACUUM runs fast, which become necessary as rows are added to or replaced in the database. I have some relatively large tables that regularly see deletes (e.g., the relational registry), and hence I want smooth vacuuming. If you don't have large tables that regularly change, you probably don't need to bother with maintenance_work_mem.

    If you have additional (or contradicting) advice on Postgres configuration for DaCHS: Please let us know, preferably on the dachs-support mailing list (see DaCHS support).

    Appendix: As I said: I was lying above. The original with-JIT plan was just fine. The horrible, cost 100 Giga, plan was only chosen when I did the SET enable_seqscan=false. Why would I do a thing like that, forcing Postgres in the wrong direction? Well, DaCHS' TAP executor makes the same setting. And why does it do that to Postgres? That's a long story closely related to the Q3C and pgsphere troubles I've mentioned above – and for which there's now finally hope: See q3c issue #30 if you're curious.

  • Semantics, Cross-Discipline Discovery, and Down-To-Earth Code

    2021-02-18 Markus Demleitner
    Boxes-and-arrows view of the UAT

    A tiny piece of the Unified Astronomy Thesaurus as viewed by Sembarebro – the IVOA logos sit on terms that have VO resoures on them.

    Sometimes people ask me (in particular when I'm wearing my hat as the current chair of the IVOA Semantics working group) “well, what's this semantics thing good for?“ There are many answers, but here's one that nicely meshes with my pet subject data discovery: You want hierarchical, agreed-upon word lists to bridge discipline gaps.

    This story starts with B2FIND, a cross-disciplinary metadata aggregator for science data run within the framework of the European Open Science Cloud (EOSC). GAVO (or, more precisely, Heidelberg University's Astronomy) is involved in the EOSC via the ESCAPE project, and so I have had the pleasure of interacting with B2FIND for a while now. In particular, they are harvesting the metadata records of the Virtual Observatory Registry from us.

    This of course requires a bit of mapping, because the VO's metadata formats (VOResource, VODataService, and several extensions; see 2014A&C.....7..101D to learn more) are far too fine-grained for the wider scientific public. Not even our good friends from high-energy physics would appreciate being served links to, say, TAP endpoints (yet!). So, on our end we're mapping to the Datacite metadata kernel, which from VOResource is just a piece of XSL away (plus some perhaps debatable conventions).

    But there's more to this mapping, such as vocabularies of subject keywords. You might argue that in the age of rapid full text searches, keywords are dead. I would beg to disagree. For example, with good, hierarchical keyword systems you can, among many other useful things, offer topical browsing of metadata repositories. While it might not quite qualify as “useful” yet, the SemBaReBro registry browser I've hacked together late last year would be an example for such facilities – and might become part of our WIRR Registry searching tool one day.

    On the topic of subject keywords VOResource says that resources in the VO should be using the Unified Astronomy Thesaurus, specifically in its IVOA incarnation (not quite true yet, but true enough by blog standards). While few do, I've done a mapping of existing keywords in the VO to UAT concepts, which is what's behind SemBaReBro. So: most VO resources now have UAT concepts.

    However, these include concepts like AM Canum Venaticorum Stars, which outside of rather specialised circles of astronomers few people will ever have heard about (which, don't get me wrong, I personally regret – they're funky star systems). Hence, B2FIND does not bother with those.

    When we discussed the subject mapping for B2FIND, we thought using the UAT's top-level concepts might be a good start. However, at that point no VO resources at all actually used these, and, indeed, within astronomy that generally wouldn't make a lot of sense, because they are to unspecific to help much within the discipline. I postponed and then forgot about the problem – when the keywords of the resources weren't even from UAT, solving the granularity mismatch just wasn't humanly possible.

    That was the state of affairs until last Tuesday, when I had a mumble session with B2FIND folks and the topic came up again. And now, thanks partly to the new desise format proposed in the current Vocabularies in the VO 2 draft, things fell nicely into place: Hey, I have UAT concepts, and mapping these to the top-level terms isn't hard either any more.

    So, B2FIND gets the toplevel keywords they've been expecting all the time starting today. Yes: This isn't a panacea suddenly solving all the problems of cross-discipline data discovery, not the least because it's harder than one might think to imagine how such a thing would look like in practice. But given the complexities involved I was positively surprised how easy this particular part of the equation was.

    From here on, there's a bit of tech babble I intend to re-use in the RFC of Vocabularies in the VO 2; don't feel bad if you skip it.

    The first step was to make the mapping from UAT terms to the toplevel terms. The interesting part of the source I'm linking to here is:

    def get_roots_for(term, uat_terms):
  roots, seen = set(), set()

  def follow(t):
    wider = uat_terms[t]["wider"]
    if not wider:
      if not t in ROOT_TERMS:
        raise Exception(
          f"{t} found as a top-level term")
      roots.add(t)
    else:
      seen.add(t)
      for wider in uat_terms[t]["wider"]:
        follow(wider)

  follow(term)
  return roots

    There, uat_terms is essentially just a json-decode of what you get from the vocabulary URI if you ask for desise (see the draft spec linked to above for the technicalities). That's really it, and it even defends against cycles in the concept graph (which are legal by SKOS but shouldn't happen in the UAT) and detached terms (i.e., ones that are not rooted in the top-level terms). For what it does, I claim that's remarkably compact code.

    Once I had that, I needed to get the UAT-mapped subject keywords for the records I'm serving to datacite and fiddle the corresponding roots back in. That's technically a bit more involved because I am producing the datacite records on the fly from the XML representation for VOResource records that I keep in the database, and there's a bit of namespace magic involved (full code). Plus, the UAT-mapped keywords are only kept in the database, not in the metadata records.

    Still, the core operation here is relatively straightforward. Consider:

    def addUATToplevels(dataciteTree):
  # dataciteTree is an (lxml) ElementTree for the
  # result of the XSL transformation.  That's all
  # I have, and thus I first have to fiddle out
  # the identifier we are talking about
  ivoid =  dataciteTree.xpath(
      "//d:alternateIdentifier["
      "@alternateIdentifierType='ivoid']",
      namespaces={"d": DATACITE_NS}
    )[0].text.lower()
  # The .lower() is necessary because ivoids
  # unfortunately are case-insensitive, and RegTAP
  # normalises them to lowercase to retain sanity.

  # Now pull the UAT-mapped subject keywords from
  # our RegTAP extension (getTableConn is
  # DaCHS-internal API, but there's no magic in
  # there, it's just connection pooling with
  # guarantees against connections  idle in
  # transaction).
  with base.getTableConn() as conn:
    subjects = set(r[0] for r in
      conn.query("SELECT uat_concept"
        " FROM rr.subject_uat"
        " WHERE ivoid=%(ivoid)s", locals()))

  # This is the mapping itself: we do
  # roots-subjects to avoid adding
  # root terms that are already in
  # the record itself.  UAT_TOPLEVELS is the result
  # of the root finding discussed above.
  for term in subjects:
    root = UAT_TOPLEVELS[term]
    newRoots |= (root-subjects)

  # And finally fiddle in any new root terms found
  # into the datacite tree
  if newRoots:
    subjects = dataciteTree.xpath(
      "//d:subjects",
      namespaces={"d": DATACITE_NS})[0]
    for root in newRoots:
      newSubject = etree.SubElement(subjects,
        f"{{{DATACITE_NS}}}subject")
      newSubject.text = root

    Apart from the technicalities I'd again say that's pretty satisfying code.

    And these two pieces of code are really all I had to do to map between the vocabularies of different granularities – which I claim will probably be the norm as metadata flows between disciplines.

    It's great to see the pieces of a fairly comples puzzle fall into place like that.

  • The Case of the Disappearing Bits

    2020-12-08 Markus Demleitner
    [number line with location markers]

    Every green line in this image stands for a value exactly representable in a floating point value of finite size. As you see, it's a white area out there [source]

    While I was preparing the publication of Coryn Bailer-Jones' distance estimations based on Gaia eDR3 (to be released about tomorrow), Coryn noticed I was swallowing digits from his numbers. My usual reaction of “aw, these are meaningless anyway because your errors are at least an order of magnitude higher” didn't work this time, because Gaia is such an incredible machine that some of the values really have six significant decimal digits. For an astronomical distance! If I had a time machine, I'd go back to F.W. Bessel right away to make him pale in envy.

    I'm storing these distances as PostgreSQL REALs, so these six digits are perilously close the seven decimal digits that the 23 bits of mantissa of single precision IEEE 754 floats are usually translated to. Suddenly, being cavalier with the last few bits of the mantissa isn't just a venial sin. It will lose science.

    So, I went hunting for the bits, going from parsing (in this case C's sscanf) through my serialisation into Postgres binary copy material (DaCHS operators: this is using a booster) to pulling the material out of the database again. And there I hit it: the bits disappeared between copying them in and retrieving them from the database.

    Wow. Turns out: It's a feature. And one I should have been aware of in that Postgres' docs have a prominent warning box where it explains its floating point types: Without setting extra-float-digits it will cut off bits. And it's done this ever since the dawn of DaCHS (in postgres terms, version 8.2 or so).

    Sure enough (edited for brevity):

    gavo=$ select r_med_geo from gedr3dist.main
gavo-$ where source_id=563018673253120;
    1430.9

gavo=$ set extra_float_digits=3;
gavo=$ select r_med_geo from gedr3dist.main
gavo-$ where source_id=563018673253120;
 1430.90332

    Starting with its database schema 26 (which is the second part of the output of dachs --version), DaCHS will configure its database roles always have extra_float_digits 3; operators beware: this may break your regression tests after the next upgrade.

    If you want to configure your non-DaCHS role, too, all it takes is:

    alter role (you) set extra_float_digits=3,

    You could also make the entire database or even the entire cluster behave like that; but then losing these bits isn't always a bad idea: It really makes the floats prettier while most of the time not losing significant data. It's just when you want to preserve the floats as you get them – and with science data, that's mostly a good idea – that we just can't really afford that prettyness.

    Update (2021-04-22): It turns out that this was already wrong (for some meaning of wrong) when I wrote this. Since PostgreSQL 12, Postgres uses shortest-precise by default (and whenever extra_float_digits is positive). The official documentation has a nice summary of the problem and the way post-12 postgres addresses it. So: expect your float-literal-comparing regression tests to break after the upgrade to bullseye.

Page 1 / 3 »