yujun777 commented on code in PR #68646:
URL: https://github.com/apache/doris/pull/68646#discussion_r4219523911
##########
fe/fe-core/src/main/java/org/apache/doris/mtmv/MTMVRelationManager.java:
##########
@@ -349,34 +360,221 @@ public void dropTable(Table table) {
// because a dropped table is the one change whose query is gone
beyond doubt. What the two record
// is the same state either way. Unlike a rename it stays an
invalidation: the table is gone for
// good, so the state is not something a later alter can make obsolete.
- processBaseTableChange(new BaseTableInfo(table), "The base table has
been deleted:", false);
+ processBaseTableChange(new BaseTableInfo(table), "The base table has
been deleted:", null);
}
/**
* update mtmv status to `SCHEMA_CHANGE`.
*
* @param isReplace
+ * @param queryJudgedColumns the names the alter gives the table or takes
away from it, which leave the
+ * judgement about each MV's state to that MV's
own query, or null when the
+ * alter is not one a query decides. The names
are carried rather than judged
+ * before the call because the judgement is
about them; see
+ * {@code AlterOp#queryJudgedColumnNames} for
which operations name one, and
+ * {@link #invalidateMvUnlessQueryHolds} for
what is asked about it. A rename
+ * of the base table names no column: it is left
to the record below, which
+ * says what the MV that keeps spelling the old
name needs to hear
*/
@Override
- public void alterTable(BaseTableInfo oldTableInfo, Optional<BaseTableInfo>
newTableInfo, boolean isReplace) {
+ public void alterTable(BaseTableInfo oldTableInfo, Optional<BaseTableInfo>
newTableInfo, boolean isReplace,
+ QueryJudgedChange queryJudgedChange) {
// when replace, need deal two table
if (isReplace) {
// REPLACE TABLE already invalidates the IVM baseline explicitly,
see Alter#processReplaceTable
- processBaseTableChange(newTableInfo.get(), "The base table has
been updated:", false);
+ processBaseTableChange(newTableInfo.get(), "The base table has
been updated:", null);
}
- boolean renamed = !isReplace && newTableInfo.isPresent()
- && !Objects.equals(oldTableInfo.getTableName(),
newTableInfo.get().getTableName());
- // A rename is the one change whose query check is skipped: the MV
query keeps spelling the old
- // name, so it is unanalyzable by construction, and the reason it
would be invalidated with --
- // "the query is no longer analyzable" -- says less than the message
this call records anyway.
- boolean checkQueryUsable = !renamed;
- processBaseTableChange(oldTableInfo, "The base table has been
updated:", checkQueryUsable);
+ processBaseTableChange(oldTableInfo, "The base table has been
updated:", queryJudgedChange);
}
/**
- * An MV's query is only as good as the base table schema it was analyzed
against. Re-analyzing the
- * MV query here (right after the alter was applied) is what detects a
changed column identity:
+ * Whether the query, as it is analysed now, reads a column of any of
these names, and reads it where
+ * the change can reach it.
+ *
+ * <p>There are two places a name is the change's to answer for. One is a
column of the table the change
+ * is about: that is the column this view's rows were computed from, and
the names are matched
+ * case-insensitively because a name is what moves. The other is a column
the query reaches across a
+ * scope boundary -- the plan records those on the Apply that stands for
the subquery, whose correlation
+ * slots are the outer columns its right side reads -- because such a name
is the scopes' to answer for
+ * rather than the query's: the nearest column to the reference answers
for it, so a column the change
+ * takes away from a scope inside leaves the name to one outside, and a
column it gives to a scope inside
+ * takes the name over. A name reached with the qualifier of another table
inside the query's own scope
+ * is neither: no later change can move it, so one to a column it does not
name is one this view's rows
+ * do not depend on.
+ */
+ private static boolean reachesAnyColumnOf(Plan plan, BaseTableInfo
baseTableInfo, Set<String> columnNames) {
+ if (plan == null) {
+ // A query whose plan was not kept is one this cannot be answered
about, and "it does" is the
+ // answer that keeps the view safe.
+ return true;
+ }
+ Set<String> names = Sets.newTreeSet(String.CASE_INSENSITIVE_ORDER);
+ names.addAll(columnNames);
+ LineageInfo lineage = LineageInfoExtractor.extractLineageInfo(plan);
+ for (SetMultimap<?, Expression> byType :
lineage.getDirectLineageMap().values()) {
+ if (reachesAnyColumn(byType.values(), names, baseTableInfo)) {
+ return true;
+ }
+ }
+ // The dataset predicates once, not once per output column: the
per-output copy of them the lineage
+ // also offers holds the same expressions for every column the query
produces, and scanning it would
+ // visit each of them once per column.
+ if (reachesAnyColumn(lineage.getDatasetIndirectLineageMap().values(),
names, baseTableInfo)) {
+ return true;
+ }
+ if (reachesAnyColumnOfASubquery(plan, names, baseTableInfo)) {
+ return true;
+ }
+ return reachesAnyColumnAcrossScopes(plan, lineage, names,
baseTableInfo);
+ }
+
+ /** Whether this slot is a column of this table, through whatever views
stand between the two. */
+ private static boolean isColumnOf(Slot slot, BaseTableInfo baseTableInfo) {
+ if (!(slot instanceof SlotReference)) {
+ return false;
+ }
+ return ((SlotReference) slot).getOriginalTable()
+ .map(table -> new BaseTableInfo(table).equals(baseTableInfo))
+ .orElse(false);
+ }
+
+ /**
+ * Whether a name the change is about is answered for inside a subquery,
out of that subquery's own
+ * scope.
+ *
+ * <p>This is the one place a name can move without any column the view
produces depending on it: the
+ * scope of a subquery is internal, so which column answers for a name
there changes what the query
+ * returns -- a row, or none -- while every column of the view stays the
one it was. The lineage of the
+ * view's columns does not reach it, so the scope the subquery became is
read here, expression by
+ * expression, the way the lineage is read for the view's own.
+ *
+ * <p>Two things are read. One is a value the subquery itself names -- an
expression of its own under one
+ * of these names, rather than a column of a table -- because that is what
a name the change takes away
+ * falls back to, and it decides the rows whether the subquery is a
predicate or a value. It is read only
+ * where the table the change is about is one that subquery reads: what a
name falls back to is what the
+ * scope that answered for it holds, so a scope that does not read the
table holds nothing for the name
+ * and one of its own is one the change never reached. The other is a
column of the table the change is
+ * about, which decides the rows only when the subquery's output is one
the query reads: an EXISTS tests
+ * the rows of its subquery and not what it projects, so a name it
projects and never compares is one
+ * this view's rows do not depend on.
+ *
+ * <p>Each scope is read on its own. A subquery inside one of these is a
scope of its own, and it is
+ * judged where it is read and not as a part of its enclosing one: its
projection is held only where
+ * that scope's own output is read, so an EXISTS inside an IN is still not
compared by the IN.
+ */
+ private static boolean reachesAnyColumnOfASubquery(Plan plan, Set<String>
names,
+ BaseTableInfo baseTableInfo) {
+ for (LogicalApply<?, ?> apply :
plan.<LogicalApply>collectToList(LogicalApply.class::isInstance)) {
+ List<Plan> itsOwnScope = ownScopeOf(apply);
+ boolean outputDecidesRows = !apply.isExist();
+ boolean isOneOfItsTables = readsAnyTableOf(itsOwnScope,
baseTableInfo);
+ for (Plan node : itsOwnScope) {
+ for (Expression expression : node.getExpressions()) {
+ if ((isOneOfItsTables &&
readsAnyNameTheSubqueryAnswersFor(expression, names))
+ || (outputDecidesRows &&
reachesAnyColumn(expression, names, baseTableInfo))) {
+ return true;
+ }
+ }
+ }
+ }
+ return false;
+ }
+
+ /**
+ * The nodes of a subquery that belong to the scope this Apply stands for:
its right side, with the
+ * right side of every subquery nested in it left out.
+ *
+ * <p>A nested Apply is a scope of its own and is read as one, so its
right side belongs to that read
+ * rather than to this one; the relation it is asked about is the
enclosing scope's own and stays.
+ */
+ private static List<Plan> ownScopeOf(LogicalApply<?, ?> apply) {
+ List<Plan> itsOwnScope = Lists.newArrayList();
+ collectItsOwnScope((Plan) apply.right(), itsOwnScope);
+ return itsOwnScope;
+ }
+
+ private static void collectItsOwnScope(Plan node, List<Plan> itsOwnScope) {
+ itsOwnScope.add(node);
+ if (node instanceof LogicalApply) {
+ collectItsOwnScope(((LogicalApply<?, ?>) node).left(),
itsOwnScope);
+ return;
+ }
+ for (Plan child : node.children()) {
+ collectItsOwnScope(child, itsOwnScope);
+ }
+ }
+
+ /** Whether one of these nodes reads this table, through whatever views
stand between the two. */
+ private static boolean readsAnyTableOf(List<Plan> itsOwnScope,
BaseTableInfo baseTableInfo) {
+ return itsOwnScope.stream().anyMatch(node -> node instanceof
LogicalCatalogRelation
+ && new BaseTableInfo(((LogicalCatalogRelation)
node).getTable()).equals(baseTableInfo));
+ }
+
+ /**
+ * Whether this expression reads a value the subquery answers for itself,
under one of these names: a
+ * slot of the subquery's own -- an alias or a value it computed -- rather
than a column of a table.
+ *
+ * <p>Read rather than merely named, because an expression of the subquery
carrying one of these names
+ * says nothing on its own: a subquery that names a `flag` of its own
while no reference in it resolves
+ * to that name is one whose rows the change cannot reach, and one that
reads the name it names is where
+ * a reference that answered for the changed column falls back to.
+ */
+ private static boolean readsAnyNameTheSubqueryAnswersFor(Expression
expression, Set<String> names) {
+ for (Slot slot : expression.getInputSlots()) {
Review Comment:
Not fixed in 5dcbddd3e14, and I would rather leave it with the reason
written down than add an exception for it.
The movement is real and I measured it on the shape in the comment: after
the light `DROP COLUMN changed_t.flag` the view stays NORMAL and in sync,
holding no rows, while the same query returns `(1)`. It is the same movement
the constant-alias shape already handles -- a name the view's query never bound
to a column, left to whatever the scope answers for it once the column is gone
-- with the difference that the value it falls back to here is an alias over
another table's column rather than a constant, and the check's test for "a
value the scope answers for itself" requires the slot to come from no table at
all, so a source-backed alias is skipped.
What keeps me from relaxing that test here is the class it would match on
the other side. The test would become "not a column of the table the change is
about", which also matches every reference in a scope that reads the changed
table whose name is another table's column -- an explicitly qualified `b.flag`
included -- and that is a wider class than the one it closes: each member costs
a dependent view a whole rebuild and its rewrite snapshot, which is what the
over-invalidation threads around this one are about. Trading a narrower missed
invalidation for a wider needless one is not a trade I can make inside this
change.
What bears on the priority: the view in this shape is not one the
transparent rewrite serves at all -- `EXPLAIN` refuses it before any change is
made (`RBO ... FailInfo: View original struct info is invalid`), while the
constant-alias shape is matched by the same stage. So the exposure left is the
view's own contents and the views built on it, rather than a query answered
from a stale view.
Both sides of this come from the same place: the check judges a name against
the plan after the change instead of comparing it with the binding the view was
built with. Closing the family there -- rather than by another clause-shaped
exception -- is what I would want for this thread.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
