Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

[bugfix] check key ordering when isSorted = true #6267

Merged
merged 4 commits into from Jun 7, 2019
Merged

[bugfix] check key ordering when isSorted = true #6267

Show file tree
Hide file tree
Changes from 2 commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
7ad53cb
check key ordering when asserting that is sorted
Jun 5, 2019
16f0f6f
Fix MatrixEntriesTable lowering to deal with duplicate row keys.
tpoterba Jun 6, 2019
ec98f40
address Patrick's comments
tpoterba Jun 6, 2019
db9ec56
bump
tpoterba Jun 7, 2019
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Jump to
Jump to file
Failed to load files.
Diff view
Diff view
6 changes: 6 additions & 0 deletions hail/python/benchmark/matrix_table_benchmarks.py
View file
Open in desktop
Expand Up @@ -20,6 +20,12 @@ def matrix_table_entries_table():
mt = hl.read_matrix_table(resource('profile.mt'))
mt.entries()._force_count()


@benchmark
def matrix_table_entries_table_no_key():
mt = hl.read_matrix_table(resource('profile.mt')).key_rows_by().key_cols_by()
mt.entries()._force_count()

@benchmark
def matrix_table_rows_force_count():
ht = hl.read_matrix_table(resource('profile.mt')).rows().key_by()
Expand Down
12 changes: 12 additions & 0 deletions hail/src/main/scala/is/hail/expr/ir/IRBuilder.scala
View file
Open in desktop
Expand Up @@ -230,6 +230,18 @@ object IRBuilder {

def dropFields(fields: Symbol*): IRProxy = dropFieldList(fields.map(_.name))

def insertStruct(other: IRProxy): IRProxy = (env: E) => {
val right = other(env)
val sym = genUID()
Let(
sym,
right,
InsertFields(
ir(env),
right.typ.asInstanceOf[TStruct].fieldNames.map(f => f -> GetField(Ref(sym, right.typ), f)),
None))
}

def len: IRProxy = (env: E) => ArrayLen(ir(env))

def isNA: IRProxy = (env: E) => IsNA(ir(env))
Expand Down
71 changes: 48 additions & 23 deletions hail/src/main/scala/is/hail/expr/ir/LowerMatrixIR.scala
View file
Open in desktop
Expand Up @@ -513,29 +513,54 @@ object LowerMatrixIR {
.rename(Map(entriesFieldName -> entries), Map(colsFieldName -> cols))

case MatrixEntriesTable(child) =>
val oldColIdx = Symbol(genUID())
val lambdaIdx1 = Symbol(genUID())
val currentColIdx = Symbol(genUID())
lower(child, ab)
.mapGlobals('global.insertFields(oldColIdx ->
irRange(0, 'global (colsField).len)
.map(lambdaIdx1 ~> makeStruct('key -> 'global (colsField)(lambdaIdx1).selectFields(child.typ.colKey: _*), 'value -> lambdaIdx1))
.sort(ascending = true, onKey = true)
.map(lambdaIdx1 ~> lambdaIdx1('value))))
.mapRows('row.insertFields(currentColIdx -> 'global (oldColIdx)
.filter(lambdaIdx1 ~> !'row (entriesField)(lambdaIdx1).isNA)))
.explode(currentColIdx)
.mapRows(let(
__current_idx = 'row (currentColIdx),
__col_struct = 'global (colsField)('__current_idx),
__entry_struct = 'row (entriesField)('__current_idx)) {
val newFields = child.typ.colType.fieldNames.map(Symbol(_)).map(f => f -> '__col_struct (f)) ++
child.typ.entryType.fieldNames.map(Symbol(_)).map(f => f -> '__entry_struct (f))
'row
.dropFields(entriesField, currentColIdx)
.insertFields(newFields: _*)
}).mapGlobals('global.dropFields(colsField, oldColIdx))
.keyBy(child.typ.rowKey ++ child.typ.colKey, isSorted = !(child.typ.rowKey.isEmpty && child.typ.colKey.nonEmpty))
val lc = lower(child, ab)

if (child.typ.colKey.nonEmpty) {
val oldColIdx = Symbol(genUID())
val lambdaIdx1 = Symbol(genUID())
val lambdaIdx2 = Symbol(genUID())
val lambdaIdx3 = Symbol(genUID())
val toExplode = Symbol(genUID())
val values = Symbol(genUID())
lc
.mapGlobals('global.insertFields(oldColIdx ->
irRange(0, 'global (colsField).len)
.map(lambdaIdx1 ~> makeStruct('key -> 'global (colsField)(lambdaIdx1).selectFields(child.typ.colKey: _*), 'value -> lambdaIdx1))
.sort(ascending = true, onKey = true)
.map(lambdaIdx1 ~> lambdaIdx1('value))))
.aggregateByKey(makeStruct(values -> applyAggOp(Collect(), seqOpArgs = FastIndexedSeq('row.selectFields(lc.typ.valueType.fieldNames: _*)))))
.mapRows('row.dropFields(values).insertFields(toExplode ->
'global (oldColIdx)
.flatMap(lambdaIdx1 ~> 'row (values)
.filter(lambdaIdx2 ~> !lambdaIdx2(entriesField)(lambdaIdx1).isNA)
.map(lambdaIdx3 ~> let(__col = 'global (colsField)(lambdaIdx1), __entry = lambdaIdx3(entriesField)(lambdaIdx1)) {
makeStruct(
child.typ.rowValueStruct.fieldNames.map(Symbol(_)).map(f => f -> lambdaIdx3(f)) ++
child.typ.colType.fieldNames.map(Symbol(_)).map(f => f -> '__col (f)) ++
child.typ.entryType.fieldNames.map(Symbol(_)).map(f => f -> '__entry (f)): _*
)
}))))

.explode(toExplode)
.mapRows('row.dropFields(toExplode).insertStruct('row (toExplode)))
.mapGlobals('global.dropFields(colsField, oldColIdx))
.keyBy(child.typ.rowKey ++ child.typ.colKey, isSorted = !(child.typ.rowKey.isEmpty && child.typ.colKey.nonEmpty))
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think you've ensured that this is always sorted. However, if the row key is empty (as uncommon as that might be), we don't want to be doing a collect-by-key (i.e., a collect). In that case I think we just have to shuffle.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

ah, yes, good point.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

ok, added that logic and fixed the field-order problem I had separately.

} else {
val colIdx = Symbol(genUID())
val lambdaIdx = Symbol(genUID())
lc
.mapRows('row.insertFields(colIdx -> irRange(0, 'global (colsField).len)
.filter(lambdaIdx ~> !'row (entriesField)(lambdaIdx).isNA)))
.explode(colIdx)
.mapRows(let(__col_struct = 'global (colsField)('row (colIdx)),
__entry_struct = 'row (entriesField)('row (colIdx))) {
val newFields = child.typ.colType.fieldNames.map(Symbol(_)).map(f => f -> '__col_struct (f)) ++
child.typ.entryType.fieldNames.map(Symbol(_)).map(f => f -> '__entry_struct (f))

'row.dropFields(entriesField, colIdx).insertFields(newFields: _*)
})
.mapGlobals('global.dropFields(colsField))
}

case MatrixToTableApply(child, function) =>
val loweredChild = lower(child, ab)
Expand Down
115 changes: 58 additions & 57 deletions hail/src/main/scala/is/hail/rvd/RVD.scala
View file
Open in desktop
Expand Up @@ -105,7 +105,7 @@ class RVD(
if (nPreservedFields == newKey.length)
this
else if (isSorted)
truncateKey(newKey.take(nPreservedFields)).extendKeyPreservesPartitioning(newKey)
truncateKey(newKey.take(nPreservedFields)).extendKeyPreservesPartitioning(newKey).checkKeyOrdering()
else
changeKey(newKey)
}
Expand All @@ -130,6 +130,61 @@ class RVD(
}
}

def checkKeyOrdering(): RVD = {
val partitionerBc = partitioner.broadcast(crdd.sparkContext)
val localType = typ
val localKPType = typ.kType

new RVD(
typ,
partitioner,
crdd.cmapPartitionsWithIndex { case (i, ctx, it) =>
val prevK = WritableRegionValue(localType.kType, ctx.freshRegion)
val kUR = new UnsafeRow(localKPType)

new Iterator[RegionValue] {
var first = true

def hasNext: Boolean = it.hasNext

def next(): RegionValue = {
val rv = it.next()

if (first)
first = false
else {
if (localType.kRowOrd.gt(prevK.value, rv)) {
kUR.set(prevK.value)
val prevKeyString = kUR.toString()

prevK.setSelect(localType.rowType, localType.kFieldIdx, rv)
kUR.set(prevK.value)
val currKeyString = kUR.toString()
fatal(
s"""RVD error! Keys found out of order:
| Current key: $currKeyString
| Previous key: $prevKeyString
|This error can occur after a split_multi if the dataset
|contains both multiallelic variants and duplicated loci.
""".stripMargin)
}
}

prevK.setSelect(localType.rowType, localType.kFieldIdx, rv)
kUR.set(prevK.value)

if (!partitionerBc.value.rangeBounds(i).contains(localType.kType.virtualType.ordering, kUR))
fatal(
s"""RVD error! Unexpected key in partition $i
| Range bounds for partition $i: ${ partitionerBc.value.rangeBounds(i) }
| Range of partition IDs for key: [${ partitionerBc.value.lowerBound(kUR) }, ${ partitionerBc.value.upperBound(kUR) })
| Invalid key: ${ kUR.toString() }""".stripMargin)
rv
}
}
})
}

def truncateKey(n: Int): RVD = {
require(n <= typ.key.length)
truncateKey(typ.key.take(n))
Expand Down Expand Up @@ -1275,62 +1330,8 @@ object RVD {
): RVD = {
if (!HailContext.get.checkRVDKeys)
return new RVD(typ, partitioner, crdd)

val sc = crdd.sparkContext

val partitionerBc = partitioner.broadcast(sc)
val localType = typ
val localKPType = typ.kType

new RVD(
typ,
partitioner,
crdd.cmapPartitionsWithIndex { case (i, ctx, it) =>
val prevK = WritableRegionValue(localType.kType, ctx.freshRegion)
val kUR = new UnsafeRow(localKPType)

new Iterator[RegionValue] {
var first = true

def hasNext: Boolean = it.hasNext

def next(): RegionValue = {
val rv = it.next()

if (first)
first = false
else {
if (localType.kRowOrd.gt(prevK.value, rv)) {
kUR.set(prevK.value)
val prevKeyString = kUR.toString()

prevK.setSelect(localType.rowType, localType.kFieldIdx, rv)
kUR.set(prevK.value)
val currKeyString = kUR.toString()
fatal(
s"""RVD error! Keys found out of order:
| Current key: $currKeyString
| Previous key: $prevKeyString
|This error can occur after a split_multi if the dataset
|contains both multiallelic variants and duplicated loci.
""".stripMargin)
}
}

prevK.setSelect(localType.rowType, localType.kFieldIdx, rv)
kUR.set(prevK.value)

if (!partitionerBc.value.rangeBounds(i).contains(localType.kType.virtualType.ordering, kUR))
fatal(
s"""RVD error! Unexpected key in partition $i
| Range bounds for partition $i: ${ partitionerBc.value.rangeBounds(i) }
| Range of partition IDs for key: [${ partitionerBc.value.lowerBound(kUR) }, ${ partitionerBc.value.upperBound(kUR) })
| Invalid key: ${ kUR.toString() }""".stripMargin)

rv
}
}
})
else
return new RVD(typ, partitioner, crdd).checkKeyOrdering()
}

def union(rvds: Seq[RVD], joinKey: Int): RVD = rvds match {
Expand Down