MSQ: Validate that strings and string arrays are not mixed. #15920

gianm · 2024-02-17T02:45:49Z

When multi-value strings and string arrays coexist in the same column, it causes problems with "classic MVD" style queries such as:

  select * from wikipedia -- fails at runtime
  select count(*) from wikipedia where flags = 'B' -- fails at planning time
  select flags, count(*) from wikipedia group by 1 -- fails at runtime

To avoid these problems, this patch adds type verification for INSERT and REPLACE. It is targeted: the only type changes that are blocked are string-to-array and array-to-string. There is also a way to exclude certain columns from the type checks, if the user really knows what they're doing.

The patch includes new documentation about arrayIngestMode; see docs/querying/arrays.md. Other mentions of it throughout the docs are updated to link to this section.

Error scenarios:

Write SQL ARRAY to existing ARRAY field in arrayIngestMode: mvd

Cannot write into field[arrayString] using type[VARCHAR ARRAY] and arrayIngestMode[mvd], since the existing type is[VARCHAR ARRAY]. Try setting arrayIngestMode to[array] to retain the SQL type[VARCHAR ARRAY]. See https://druid.apache.org/docs/latest/querying/arrays#arrayingestmode for more details about this check and how to override it if needed.

Write SQL VARCHAR to existing ARRAY field in arrayIngestMode: mvd

Cannot write into field[arrayString] using type[VARCHAR] and arrayIngestMode[mvd], since the existing type is[VARCHAR ARRAY]. Try setting arrayIngestMode to[array] and adjusting your query to make this column an ARRAY instead of VARCHAR. See https://druid.apache.org/docs/latest/querying/arrays#arrayingestmode for more details about this check and how to override it if needed.

Write SQL VARCHAR to existing ARRAY field in arrayIngestMode: array

Cannot write into field[arrayString] using type[VARCHAR] and arrayIngestMode[array], since the existing type is[VARCHAR ARRAY]. Try adjusting your query to make this column an ARRAY instead of VARCHAR. See https://druid.apache.org/docs/latest/querying/arrays#arrayingestmode for more details about this check and how to override it if needed.

Write SQL ARRAY to existing VARCHAR field in arrayIngestMode: array

Cannot write into field[mvString] using type[VARCHAR ARRAY] and arrayIngestMode[array], since the existing type is[VARCHAR]. Try wrapping this field using ARRAY_TO_MV(...) AS "mvString". See https://druid.apache.org/docs/latest/querying/arrays#arrayingestmode for more details about this check and how to override it if needed.

When multi-value strings and string arrays coexist in the same column, it causes problems with "classic MVD" style queries such as: select * from wikipedia -- fails at runtime select count(*) from wikipedia where flags = 'B' -- fails at planning time select flags, count(*) from wikipedia group by 1 -- fails at runtime To avoid these problems, this patch adds type verification for INSERT and REPLACE. It is targeted: the only type changes that are blocked are string-to-array and array-to-string. There is also a way to exclude certain columns from the type checks, if the user really knows what they're doing.

gianm · 2024-02-17T02:46:07Z

This patch does not yet include docs and tests, so marking it draft. The idea is there though.

gianm · 2024-02-17T02:55:19Z

Depending on which direction you're going in, the errors look like either:

Cannot write into field[flags] using type[VARCHAR ARRAY] and arrayIngestMode[mvd], since the existing type is[VARCHAR ARRAY]. Try setting arrayIngestMode to[array] to retain the SQL type[VARCHAR ARRAY]

Or:

Cannot write into field[flags] using type[VARCHAR ARRAY] and arrayIngestMode[array], since the existing type is[VARCHAR]. Try wrapping this field using ARRAY_TO_MV(...) AS "flags"

clintropolis

👍

clintropolis · 2024-02-17T22:36:55Z

extensions-core/multi-stage-query/src/main/java/org/apache/druid/msq/sql/MSQTaskSqlEngine.java

+  )
+  {
+    if (targetTable == null) {
+      return;


I suppose this doesn't necessarily need to be done as part of this PR, but I do want to see at some point that this would have a validation failure for the case where the output type has any ARRAY and arrayIngestMode isn't set to array, for the effect of preventing people from writing new insert queries that are incorrectly relying on the mvd behavior.

I was thinking more about the validations that are in effect when arrayIngestMode: array, since those validations would help us change the default.

Maybe when we change the default, we can also consider making it so when people explicitly set arrayIngestMode: mvd, they need to use ARRAY_TO_MV. (I think that's what you're suggesting here.)

Maybe when we change the default, we can also consider making it so when people explicitly set arrayIngestMode: mvd, they need to use ARRAY_TO_MV. (I think that's what you're suggesting here.)

yea, I'm mostly thinking about it in terms of forcing people to write the correct queries as soon as possible in either mode to make things less painful down the road, and was thinking it would be nice to prevent bad habits for new tables before they have a chance to get started since those habits are going to have to change.

Its ok to wait until we change the default i guess, which this change should allow us to do much sooner (druid 30 hopefully)

LakshSingla · 2024-02-18T16:25:03Z

...s-core/multi-stage-query/src/main/java/org/apache/druid/msq/util/MultiStageQueryContext.java

@@ -152,6 +154,7 @@ public class MultiStageQueryContext
  public static final String CTX_ARRAY_INGEST_MODE = "arrayIngestMode";
  public static final ArrayIngestMode DEFAULT_ARRAY_INGEST_MODE = ArrayIngestMode.MVD;

+  public static final String CTX_SKIP_TYPE_VERIFICATION = "skipTypeVerification";


Would we have docs for this parameter? If so, it'll be another parameter that we'd have to support in MSQ. If not, then I don't see a way for users to set this parameter, both of which seem ugly.

Can we do without this parameter, since the user can use ARRAY_TO_MV or MV_TO_ARRAY functions to preserve the behaviour?

The purpose of this parameter is to allow someone to explicitly decide to ingest mismatched types. I don't think ARRAY_TO_MV and MV_TO_ARRAY would be helpful, since they're more something that users could use to make types line up properly.

One scenario for this parameter might be that someone is migrating from MVD to array, and has updated their queries to work with mismatched array and MVD types: they added MV_TO_ARRAY around all their MVDs, and have migrated to UNNEST, etc. They want to start ingesting arrays rather than MVDs for new data, but don't want to go back and rewrite all their old data, because that would be too expensive.

I do agree it's not really ideal for it to be a new context parameter. I just couldn't think of another approach to allowing validation bypass for specific columns, without a table-level catalog.

Maybe some new SQL syntax?

I wonder if you have any thoughts here? (or anyone else?)

the context flag seems ok to me, especially since i don't think we're quite yet in a place where a lot of people will want to be using a mix of mvds and arrays, we are missing optimized versions of MV_TO_ARRAY and ARRAY_TO_MV to make things a lot more efficient to use a mix of stuff.

I feel like once we've transitioned to array being default arrayIngestMode, given some time for it to settle, and then eventually remove arrayIngestMode to drop all of the other behaviors, we might be able to consider removing this validation in favor of pushing it to the catalog since the old habits would be gone and at that point it should also be a lot less painful to work with a mix of stuff. But we can revisit once arrayIngestMode is removed.

LakshSingla · 2024-02-18T16:34:32Z

extensions-core/multi-stage-query/src/main/java/org/apache/druid/msq/sql/MSQTaskSqlEngine.java

+        final ColumnType newDruidType =
+            DimensionSchemaUtils.getDimensionType(Calcites.getColumnTypeForRelDataType(newSqlType), arrayIngestMode);
+
+        if (newDruidType.isArray() && oldDruidType.is(ValueType.STRING)


This seems like we are building a DDL layer, but for specific types, and for MSQ ingestion, which seems kinda off, given that Druid allows a flexible schema. However, I also think that most of the cases triggering this would be accidental, rather than deliberate. Do we plan on getting rid of these validations, once arrayIngestMode = array is the only acceptable argument to MSQ ingestion, and we are not stuck between supporting incorrect behavior, and allowing only correct but incompatible behavior?

This validation is really just to prevent accidents that break queries, due a pretty unique situation:

due to the complexity around arrayIngestMode it is easy to accidentally ingest a list of strings as the "wrong" type accidentally (MVD when you meant ARRAY or vice-versa).

when a single column has both MVD and ARRAY in it, existing queries can break for the reasons I mentioned in the PR description.

This situation is problematic but also only really exists for the MVD/array combinaton. So I don't intent this to be the start of a DDL layer, just a point fix to this particular problem. I do think we want a catalog layer that can be part of more extensive type validations, but that should be opt-in & user-driven & more flexible than some hard-coded checks.

Thanks for the explanation!

gianm · 2024-03-13T06:10:32Z

Taking this out of draft since I've added tests and docs. I have also added error message examples to the PR description.

clintropolis · 2024-03-13T06:20:03Z

docs/querying/arrays.md

+This causes them to be stored as [multi-value strings](multi-value-dimensions.md), using the same `STRING` column type
+as regular scalar strings. SQL `BIGINT ARRAY` and `DOUBLE ARRAY` cannot be loaded under `arrayIngestMode: mvd`. This
+is the default behavior when `arrayIngestMode` is not provided in your query context, although the default behavior
+may change to `array` in a future release.


nit: will change

clintropolis · 2024-03-13T06:21:33Z

docs/querying/arrays.md

+  ) EXTEND (
+    "timestamp" VARCHAR,
+    "label" VARCHAR,
+    "arrayString" VARCHAR ARRAY,
+    "arrayLong" BIGINT ARRAY,
+    "arrayDouble" DOUBLE ARRAY


thanks for fixing these up, i wrote these docs before i fixed up EXTEND to support array types 🤘

clintropolis · 2024-03-13T06:27:32Z

extensions-core/multi-stage-query/src/main/java/org/apache/druid/msq/sql/MSQTaskSqlEngine.java

+  )
+  {
+    if (targetTable == null) {
+      return;


Maybe when we change the default, we can also consider making it so when people explicitly set arrayIngestMode: mvd, they need to use ARRAY_TO_MV. (I think that's what you're suggesting here.)

yea, I'm mostly thinking about it in terms of forcing people to write the correct queries as soon as possible in either mode to make things less painful down the road, and was thinking it would be nice to prevent bad habits for new tables before they have a chance to get started since those habits are going to have to change.

Its ok to wait until we change the default i guess, which this change should allow us to do much sooner (druid 30 hopefully)

clintropolis · 2024-03-13T06:30:11Z

...ons-core/multi-stage-query/src/main/java/org/apache/druid/msq/util/DimensionSchemaUtils.java

+  {
+    if (queryType == null) {
+      // if schema information is not available, create a string dimension
+      return ColumnType.STRING;


i wonder if this should return null to make createDimensionSchema decided to use the string schema (in case down the line we ever want to switch that to the auto schema)

I held this condition over from the previous code, but I'm not sure that the type can actually ever be null. This method is used in a couple places:

by MSQ ControllerImpl, which checks for null before passing it; if it's null a "No type for column" exception is thrown.

by the new validation code I added; but here null only arises if Calcites.getColumnTypeForRelDataType(newSqlType) is null, which seems unlikely for a SQL type that appears in an already-planned INSERT or REPLACE statement. The fact that it's already been planned means we should have Druid types corresponding to all the SQL types.

So, maybe we could remove the @Nullable and make this an error.

Should I try that now, or want to leave it for later?

ah, I should have looked at the callers, that makes sense, leaving for now and making an error later seems fine to me

vogievetsky

Just built and plaid with it. Looks great.

…5920) * MSQ: Validate that strings and string arrays are not mixed. When multi-value strings and string arrays coexist in the same column, it causes problems with "classic MVD" style queries such as: select * from wikipedia -- fails at runtime select count(*) from wikipedia where flags = 'B' -- fails at planning time select flags, count(*) from wikipedia group by 1 -- fails at runtime To avoid these problems, this patch adds type verification for INSERT and REPLACE. It is targeted: the only type changes that are blocked are string-to-array and array-to-string. There is also a way to exclude certain columns from the type checks, if the user really knows what they're doing. * Fixes. * Tests and docs and error messages. * More docs. * Adjustments. * Adjust message. * Fix tests. * Fix test in DV mode.

#15920) (#16160) * Cherry-picking 15920-to-29.0.1 * Fixing extra test case which got added as part of merge --------- Co-authored-by: Gian Merlino <[email protected]>

github-actions bot added Area - Batch Ingestion Area - Querying Area - MSQ For multi stage queries - https://github.com/apache/druid/issues/12262 labels Feb 17, 2024

gianm marked this pull request as draft February 17, 2024 02:46

Fixes.

083f736

clintropolis reviewed Feb 17, 2024

View reviewed changes

LakshSingla reviewed Feb 18, 2024

View reviewed changes

cryptoe added this to the 29.0.1 milestone Feb 21, 2024

gianm added 3 commits March 1, 2024 11:12

Merge branch 'master' into sql-ingest-validate-string-array-mismatch

092cb75

Merge branch 'master' into sql-ingest-validate-string-array-mismatch

4e01ce4

Tests and docs and error messages.

edf2843

github-actions bot added the Area - Documentation label Mar 13, 2024

gianm marked this pull request as ready for review March 13, 2024 05:49

gianm added 2 commits March 12, 2024 22:54

More docs.

09be4b6

Adjustments.

9f8f827

gianm added 2 commits March 12, 2024 23:11

Adjust message.

ef7e081

Fix tests.

a765412

clintropolis approved these changes Mar 13, 2024

View reviewed changes

Fix test in DV mode.

09d63d2

vogievetsky approved these changes Mar 13, 2024

View reviewed changes

gianm merged commit 256160a into apache:master Mar 13, 2024
83 checks passed

kgyrtkirk mentioned this pull request Mar 14, 2024

fix compile error: ARRAYS_DATASOURCE #16120

Merged

cryptoe mentioned this pull request Mar 19, 2024

[Backport] MSQ: Validate that strings and string arrays are not mixed. (#15920) #16160

Merged

cryptoe mentioned this pull request Mar 21, 2024

[Draft] 29.0.1 Release Notes #16183

Closed

clintropolis mentioned this pull request Jul 24, 2024

change arrayIngestMode default to array #16789

Merged

5 tasks

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

MSQ: Validate that strings and string arrays are not mixed. #15920

MSQ: Validate that strings and string arrays are not mixed. #15920

gianm commented Feb 17, 2024 •

edited

Loading

gianm commented Feb 17, 2024

gianm commented Feb 17, 2024

clintropolis left a comment

clintropolis Feb 17, 2024

gianm Mar 1, 2024

clintropolis Mar 13, 2024

LakshSingla Feb 18, 2024 •

edited

Loading

LakshSingla Feb 18, 2024

gianm Mar 1, 2024 •

edited

Loading

clintropolis Mar 1, 2024

LakshSingla Feb 18, 2024 •

edited

Loading

gianm Mar 1, 2024

LakshSingla Mar 6, 2024

gianm commented Mar 13, 2024

clintropolis Mar 13, 2024

clintropolis Mar 13, 2024

clintropolis Mar 13, 2024

clintropolis Mar 13, 2024

gianm Mar 13, 2024 •

edited

Loading

clintropolis Mar 13, 2024

vogievetsky left a comment

MSQ: Validate that strings and string arrays are not mixed. #15920

MSQ: Validate that strings and string arrays are not mixed. #15920

Conversation

gianm commented Feb 17, 2024 • edited Loading

gianm commented Feb 17, 2024

gianm commented Feb 17, 2024

clintropolis left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

LakshSingla Feb 18, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

gianm Mar 1, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

LakshSingla Feb 18, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

gianm commented Mar 13, 2024

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

gianm Mar 13, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

vogievetsky left a comment

Choose a reason for hiding this comment

gianm commented Feb 17, 2024 •

edited

Loading

LakshSingla Feb 18, 2024 •

edited

Loading

gianm Mar 1, 2024 •

edited

Loading

LakshSingla Feb 18, 2024 •

edited

Loading

gianm Mar 13, 2024 •

edited

Loading