Note
This technote is a work-in-progress.
1 Abstract¶
Currently we generate ObsCore tables from Registry data with a special application which uses Registry API to discover the data. This application needs to be run regularly to reflect Registry updates, which represent additional step in making ObsCore data available to users. This note discusses alternative ways to extract ObsCore data from Registry that do not require running export application.
2 ObsCore export script¶
Current procedure for export of Registry data to ObsCore format is implemented in a dax_obscore package as a stand-alone application that reads Registry and generates CSV or Parquet files. At the very high level the export application performs following operations:
It is configured with a list of collections and dataset types to read from Registry.
For each dataset type it calls
Registry.queryDatasets(dataset_type_name, collections, ...)
.Returned datasets are expanded to include all dimension records.
From the expanded
DataId
the script builds a corresponding ObsCore record. There are some non-trivial transformations performed during this stage:Some of the ObsCore columns are filled from the configuration, either global, per-dataset, or per-filter band.
As the Registry does not define per-exposure regions, the script extracts exposure region from a matching visit (or visit+detector).
Registry stores
sphgeom
regions in an opaque byte-encoded format, for ObsCore the script transforms regions into an ADQLPOLYGON
string representation.
Additional post-processing is performed when the generated data is ingested into QServ/MySQL. To accelerate point-in-region searches in the ObsCore table, a new spatially-indexed column is added to the table which defines a bounding box for a region.
2.1 Export performance¶
The Registry.queryDatasets
method performs a non-trivial queries which has performance implications.
As an example below are some numbers obtained with a recent export of DP0.2
Registry at data-int.lsst.cloud
.
Configuration included several dataset types and a chained collection “2.2i/runs/DP0.2”, the collection resolves to 286 run-type collections.
Dataset type |
Export time |
Record count |
---|---|---|
raw |
15m30s |
2,870,807 |
calexp |
16m30s |
2,805,017 |
deepCoadd_calexp |
14s |
46,158 |
goodSeeingCoadd |
11s |
46,158 |
goodSeeingDiff_differenceExp |
19m05s |
2,707,834 |
Total |
51m30s |
Approximately 90% of the total time is spent on client side CPU, this is the price to pay for iterating over millions of result rows and performing non-trivial operations on each row.
These results were obtained with a findFirst
parameter of Registry.queryDatasets
set to False
.
A naturally correct value for that parameter should be True
, which would avoid possible duplicates in the output.
Unfortunately, setting findFirst
to True
increases export time dramatically, export takes approximately 36 hours to finish with that option.
2.2 SQL queries¶
Each call to a Registry.queryDatasets
method executes a small number of SQL queries - one query to find all dataset references and followup queries to expand references with dimension records.
If the input collection has CHAINED type, then the queries need to filter all included RUN-type collections which can grow to a very long list.
Here is an example of a query for finding references calexp
dataset type with findFirst=False
(collection list is truncated, it has 58 run collections in total):
SELECT
physical_filter.band AS band,
physical_filter.instrument AS instrument,
calexp.detector AS detector,
physical_filter.name AS physical_filter,
visit.visit_system AS visit_system,
visit.id AS visit,
calexp.id AS dataset_id,
calexp.run_name AS run_name
FROM
(
SELECT
dataset_tags_00000007.instrument AS instrument,
dataset_tags_00000007.detector AS detector,
dataset_tags_00000007.visit AS visit,
dataset_tags_00000007.dataset_id AS id,
dataset.run_name AS run_name,
dataset.ingest_date AS ingest_date
FROM
dataset_tags_00000007
JOIN dataset ON dataset_tags_00000007.dataset_id = dataset.id
WHERE
dataset_tags_00000007.dataset_type_id = 41
AND dataset_tags_00000007.collection_name IN (
'2.2i/runs/DP0.2/v23_0_0_rc5/PREOPS-905/20220112T143133Z',
'2.2i/runs/DP0.2/v23_0_0_rc5/PREOPS-905/20220112T114604Z',
-- 55 other collection names
'2.2i/runs/DP0.2/v23_0_0_rc5/PREOPS-905/20211218T002844Z'
)
AND dataset.dataset_type_id = 41
) AS calexp
JOIN visit ON calexp.instrument = visit.instrument
AND calexp.visit = visit.id
JOIN physical_filter ON calexp.instrument = physical_filter.instrument
AND visit.instrument = physical_filter.instrument
AND visit.physical_filter = physical_filter.name;
When findFirst
is set to True
the queries are more complicated as they need to return a unique references based on collection ordering.
Here is an example of a query with findFirst=False
with a large fraction of sub-queries truncated:
WITH calexp_search AS (
SELECT
dataset_tags_00000007.instrument AS instrument,
dataset_tags_00000007.detector AS detector,
dataset_tags_00000007.visit AS visit,
dataset_tags_00000007.dataset_id AS id,
'2.2i/runs/DP0.2/v23_0_0_rc5/PREOPS-905/20220112T143133Z' AS run_name,
dataset.ingest_date AS ingest_date,
0 AS rank
FROM
dataset_tags_00000007
JOIN dataset ON dataset_tags_00000007.dataset_id = dataset.id
WHERE
dataset_tags_00000007.dataset_type_id = 41
AND dataset_tags_00000007.collection_name = '2.2i/runs/DP0.2/v23_0_0_rc5/PREOPS-905/20220112T143133Z'
AND dataset.dataset_type_id = 41
UNION
SELECT
dataset_tags_00000007.instrument AS instrument,
dataset_tags_00000007.detector AS detector,
dataset_tags_00000007.visit AS visit,
dataset_tags_00000007.dataset_id AS id,
'2.2i/runs/DP0.2/v23_0_0_rc5/PREOPS-905/20220112T114604Z' AS run_name,
dataset.ingest_date AS ingest_date,
1 AS rank
FROM
dataset_tags_00000007
JOIN dataset ON dataset_tags_00000007.dataset_id = dataset.id
WHERE
dataset_tags_00000007.dataset_type_id = 41
AND dataset_tags_00000007.collection_name = '2.2i/runs/DP0.2/v23_0_0_rc5/PREOPS-905/20220112T114604Z'
AND dataset.dataset_type_id = 41
UNION
-- plus 56 other queries
)
SELECT
physical_filter.band AS band,
physical_filter.instrument AS instrument,
calexp.detector AS detector,
physical_filter.name AS physical_filter,
visit.visit_system AS visit_system,
visit.id AS visit,
calexp.id AS dataset_id,
calexp.run_name AS run_name
FROM
(
SELECT
calexp_window.id AS id,
calexp_window.run_name AS run_name,
calexp_window.ingest_date AS ingest_date,
calexp_window.instrument AS instrument,
calexp_window.detector AS detector,
calexp_window.visit AS visit,
calexp_window.rownum AS rownum
FROM
(
SELECT
calexp_search.id AS id,
calexp_search.run_name AS run_name,
calexp_search.ingest_date AS ingest_date,
calexp_search.instrument AS instrument,
calexp_search.detector AS detector,
calexp_search.visit AS visit,
row_number() OVER (
PARTITION BY calexp_search.instrument,
calexp_search.detector,
calexp_search.visit
ORDER BY
calexp_search.rank
) AS rownum
FROM
calexp_search
) AS calexp_window
WHERE
calexp_window.rownum = 1
) AS calexp
JOIN visit ON calexp.instrument = visit.instrument
AND calexp.visit = visit.id
JOIN physical_filter ON calexp.instrument = physical_filter.instrument
AND visit.instrument = physical_filter.instrument
AND visit.physical_filter = physical_filter.name;
3 Ideas for alternative implementations¶
There may be few options for supporting ObsCore functionality that do not rely on export via Registry.queryDatasets
method.
These options and their related issues are discussed in sections below.
3.1 Online view of Registry tables¶
A natural idea for avoiding regular manual exports of the data is to create a SQL view that will execute the same set of queries as the export script. It should be possible to join the query from the find-first SQL listing with the additional queries to extract dimension records and represent the result as ObsCore schema. There are several issues with this approach though:
The performance of the queries. To be usable for interactive work the queries in the view need to be executed very quick, which is not the case for a correct
find-first
approach.Chained collection management. The list of the run collections in a chained collection will be growing and view needs to be aware of that. Either the view needs to be redefined each time a new run collection is added, or view has to be made smarter to use build list of run collections at execution time. It may be possible to implement latter case at the expense of even more complicated queries.
Spatial transformations. To represent region data in ObsCore ADQL format the view needs to decode binary
sphgeom
format and translate it into a different representation. This may be very non-trivial to implement in pure SQL.Indexing. Queries on ObsCore data will likely use restrictions on some columns. Without proper indexing these queries may be inefficient. Queries on view could in some cases use indices on underlying tables, but with the very complex queries there is no guarantee the query optimizer can do it in our case.
3.2 Materialized views¶
PostgreSQL supports materialized views which is a real table containing snapshot of the view query. The materialized vew is refreshed regularly bu running a special ‘REFRESH MATERIALIZED VIEW’ command which re-populates the snapshot with the current data. Because a materialized view is backed by an actual table, the indices can be defined on that table and that should help with the query performance on the view.
The issues with the performance of the view queries still has an impact with the materialized view. The REFRESH command needs to run the whole set ot queries which presently takes very long time, so the updates cannot happen frequently. Out-of-box PostgreSQL does not support incremental updates, though there is a work in progress to add incremental vew update support in the future. Even if incremental update support is added, it is very doubtful that it can handle queries of the complexity that we define for our data.
Just as with the online views, collection management is also an issue with the materialized views. Similarly the decoding of region data is an issue for any pure-SQL-based solution.
3.3 ObsCore table with triggers¶
One way to avoid running expensive queries on Registry is to insert records into an actual ObsCore table at the same time as populating regular Registry tables. At the database level this can be implemented with INSERT/DELETE triggers on relevant Registry tables. In PostgreSQL triggers can be implemented in a number of languages, including PL/Python.
Triggers have their own set of issues that make their use non-trivial:
PL/Python is implemented as a separate extension that needs to to be installed in each database.
Server-side environment could be more restricted, for example, to import
sphgeom
package it may need to be installed separately on server host.Handling of collection ordering may require non-trivial logic and queries. Although with some assumptions the ordering check may not be necessary during INSERT.
Performance needs special consideration, calling additional potentially complex code for each inserted row can make things slower.
Error and exception handling in server-side functions can be more challenging.
3.4 ObsCore table with client-side updates¶
A different INSERT-time approach could be implemented with client-side hooks. The Registry can be extended to update an additional ObsCore table while doing regular updates to its own schema. As usual, there are benefits and drawbacks in this approach:
ObsCore can be thought of as a part of Registry schema which can be covered by the existing modular “managers” design, extending that design to include ObsCore could be a natural extension.
ObsCore schema could use the same schema migration tools as other Registry managers.
ObsCore schema may be less stable, at least initially, compared to the rest of the Registry schema, which could require more frequent schema migrations.
There could be multiple ObsCore schemas in the same Registry, e.g. it may make sense to make separate ObsCore tables for each instrument.
There could also be a combined approach which stores stable parts of ObsCore schema in the table, and implements other less stable parts of the schema as a view on that table. It could also support multiple views of the same table, e.g. to satisfy specific per-instrument requirements. This approach could minimize the need for schema migrations of the actual table, replacing them with the re-definition of the views.
4 Design decisions¶
The long running time for the find-first
queries is certainly a blocker for view based implementations.
Still those solutions could be considered if query time can be drastically reduced.
Online view is probably the least favorite solution because of lack of indexing which can harm the performance of ObsCore queries.
In general, solutions based on server-side implementation, either in plain SQL or PL/Python, has potential for duplicating some of the daf_butler
complex logic.
Proving or testing that duplicated logic is implemented correctly may be a major issue.
The need to transform spatial regions from sphgeom
encoding to ADQL also makes SQL-based solutions problematic, although this issue can be mitigated by replacing sphgeom
binary encoding with some other text-based representation in the Registry schema.
4.1 Ideas for client-side implementation¶
Overall, a solution based on client-side updates may have fewer potential issues compared to other solutions. Implementation of this approach can be straightforward, though the concerns related to collection management also need a consideration. Like other parts of the Registry, the ObsCore-related operations can be implemented via a separate manager class, which can be optional for repositories that do not need ObsCore support.
There are several Registry operations that may need to update ObsCore table:
bulk import of the exported data, or ingest of raw datasets;
some regular
Butler.put()
inserts;addition of a non-empty run collection to a chained collection if chained collection is mirrored to the ObsCore table;
removal of individual datasets or a whole run collection;
removal of a run collection from a chained collection, keeping run collection.
These operation could call methods of a new ObsCore manager instance and pass the information necessary for updating ObsCore table.
Addition of the new datasets should be straightforward, minor concern is potential performance penalty to query the Registry for additional information that needs to be stored with the new dataset.
Complete removal of a dataset from the Registry should also cause its removal from ObsCore table, this could be handled transparently via the foreign key on dataset
table and CASCADE action.
Changes in collections composition may need some attention but should not be hard to implement.
The schema of the ObsCore table is likely to change over time. The configuration which controls the ObsCore export process determines both the schema of the table and its contents. To support changing configuration it should be possible to utilize the same migration tools that are used by other Registry managers. One potentially interesting issue for this approach is whether one Registry could have more than one ObsCore table based on different configurations.