feat(inverted_index): introduce SstIndexCreator (#3107)

* feat(inverted_index): introduce SstIndexCreator Signed-off-by: Zhenchi <[email protected]> * chore: tiny polish Signed-off-by: Zhenchi <[email protected]> * feat: distinguish intermediate store and index store Signed-off-by: Zhenchi <[email protected]> * chore: move comment as doc comment Signed-off-by: Zhenchi <[email protected]> * refactor: column id as index name Signed-off-by: Zhenchi <[email protected]> --------- Signed-off-by: Zhenchi <[email protected]>
GreptimeTeam · Jan 9, 2024 · db98484 · db98484
1 parent 7d0d216
commit db98484
Show file tree

Hide file tree

Showing 19 changed files with 911 additions and 47 deletions.
diff --git a/src/index/src/inverted_index/create.rs b/src/index/src/inverted_index/create.rs
@@ -12,8 +12,8 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-mod sort;
-mod sort_create;
+pub mod sort;
+pub mod sort_create;
 
 use async_trait::async_trait;
 

diff --git a/src/index/src/inverted_index/create/sort.rs b/src/index/src/inverted_index/create/sort.rs
@@ -12,8 +12,8 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-mod external_provider;
-mod external_sort;
+pub mod external_provider;
+pub mod external_sort;
 mod intermediate_rw;
 mod merge_stream;
 

diff --git a/src/index/src/inverted_index/error.rs b/src/index/src/inverted_index/error.rs
@@ -15,7 +15,7 @@
 use std::any::Any;
 use std::io::Error as IoError;
 
-use common_error::ext::ErrorExt;
+use common_error::ext::{BoxedError, ErrorExt};
 use common_error::status_code::StatusCode;
 use common_macro::stack_trace_debug;
 use snafu::{Location, Snafu};
@@ -167,6 +167,12 @@ pub enum Error {
         total_row_count: usize,
         expected_row_count: usize,
     },
+
+    #[snafu(display("External error"))]
+    External {
+        source: BoxedError,
+        location: Location,
+    },
 }
 
 impl ErrorExt for Error {
@@ -197,6 +203,8 @@ impl ErrorExt for Error {
             | FstInsert { .. }
             | InconsistentRowCount { .. }
             | IndexNotFound { .. } => StatusCode::InvalidArguments,
+
+            External { source, .. } => source.status_code(),
         }
     }
 

diff --git a/src/mito2/src/error.rs b/src/mito2/src/error.rs
@@ -445,6 +445,21 @@ pub enum Error {
         location: Location,
     },
 
+    #[snafu(display("Failed to push index value"))]
+    PushIndexValue {
+        source: index::inverted_index::error::Error,
+        location: Location,
+    },
+
+    #[snafu(display("Failed to write index completely"))]
+    IndexFinish {
+        source: index::inverted_index::error::Error,
+        location: Location,
+    },
+
+    #[snafu(display("Operate on aborted index"))]
+    OperateAbortedIndex { location: Location },
+
     #[snafu(display("Failed to read puffin metadata"))]
     PuffinReadMetadata {
         source: puffin::error::Error,
@@ -463,6 +478,18 @@ pub enum Error {
         location: Location,
     },
 
+    #[snafu(display("Failed to write puffin completely"))]
+    PuffinFinish {
+        source: puffin::error::Error,
+        location: Location,
+    },
+
+    #[snafu(display("Failed to add blob to puffin file"))]
+    PuffinAddBlob {
+        source: puffin::error::Error,
+        location: Location,
+    },
+
     #[snafu(display("Failed to clean dir {dir}"))]
     CleanDir {
         dir: String,
@@ -522,6 +549,7 @@ impl ErrorExt for Error {
             | RegionCorrupted { .. }
             | CreateDefault { .. }
             | InvalidParquet { .. }
+            | OperateAbortedIndex { .. }
             | PuffinBlobTypeNotFound { .. }
             | UnexpectedReplay { .. } => StatusCode::Unexpected,
             RegionNotFound { .. } => StatusCode::RegionNotFound,
@@ -569,10 +597,14 @@ impl ErrorExt for Error {
             EmptyRegionDir { .. } | EmptyManifestDir { .. } => StatusCode::RegionNotFound,
             ArrowReader { .. } => StatusCode::StorageUnavailable,
             ConvertValue { source, .. } => source.status_code(),
-            BuildIndexApplier { source, .. } | ApplyIndex { source, .. } => source.status_code(),
-            PuffinReadMetadata { source, .. } | PuffinReadBlob { source, .. } => {
-                source.status_code()
-            }
+            BuildIndexApplier { source, .. }
+            | PushIndexValue { source, .. }
+            | ApplyIndex { source, .. }
+            | IndexFinish { source, .. } => source.status_code(),
+            PuffinReadMetadata { source, .. }
+            | PuffinReadBlob { source, .. }
+            | PuffinFinish { source, .. }
+            | PuffinAddBlob { source, .. } => source.status_code(),
             CleanDir { .. } => StatusCode::Unexpected,
             InvalidConfig { .. } => StatusCode::InvalidArguments,
             StaleLogEntry { .. } => StatusCode::Unexpected,

diff --git a/src/mito2/src/metrics.rs b/src/mito2/src/metrics.rs
@@ -150,7 +150,7 @@ lazy_static! {
     // Index metrics.
     /// Timer of index application.
     pub static ref INDEX_APPLY_ELAPSED: Histogram = register_histogram!(
-        "index_apply_elapsed",
+        "greptime_index_apply_elapsed",
         "index apply elapsed",
     )
     .unwrap();
@@ -160,6 +160,26 @@ lazy_static! {
         "index apply memory usage",
     )
     .unwrap();
+    /// Timer of index creation.
+    pub static ref INDEX_CREATE_ELAPSED: HistogramVec = register_histogram_vec!(
+        "greptime_index_create_elapsed",
+        "index create elapsed",
+        &[STAGE_LABEL]
+    )
+    .unwrap();
+    /// Counter of rows indexed.
+    pub static ref INDEX_CREATE_ROWS_TOTAL: IntCounter = register_int_counter!(
+        "greptime_index_create_rows_total",
+        "index create rows total",
+    )
+    .unwrap();
+    /// Counter of created index bytes.
+    pub static ref INDEX_CREATE_BYTES_TOTAL: IntCounter = register_int_counter!(
+        "greptime_index_create_bytes_total",
+        "index create bytes total",
+    )
+    .unwrap();
+
     /// Counter of r/w bytes on index related IO operations.
     pub static ref INDEX_IO_BYTES_TOTAL: IntCounterVec = register_int_counter_vec!(
         "greptime_index_io_bytes_total",
@@ -170,6 +190,15 @@ lazy_static! {
     /// Counter of read bytes on puffin files.
     pub static ref INDEX_PUFFIN_READ_BYTES_TOTAL: IntCounter = INDEX_IO_BYTES_TOTAL
         .with_label_values(&["read", "puffin"]);
+    /// Counter of write bytes on puffin files.
+    pub static ref INDEX_PUFFIN_WRITE_BYTES_TOTAL: IntCounter = INDEX_IO_BYTES_TOTAL
+        .with_label_values(&["write", "puffin"]);
+    /// Counter of read bytes on intermediate files.
+    pub static ref INDEX_INTERMEDIATE_READ_BYTES_TOTAL: IntCounter = INDEX_IO_BYTES_TOTAL
+        .with_label_values(&["read", "intermediate"]);
+    /// Counter of write bytes on intermediate files.
+    pub static ref INDEX_INTERMEDIATE_WRITE_BYTES_TOTAL: IntCounter = INDEX_IO_BYTES_TOTAL
+        .with_label_values(&["write", "intermediate"]);
 
     /// Counter of r/w operations on index related IO operations, e.g. read, write, seek and flush.
     pub static ref INDEX_IO_OP_TOTAL: IntCounterVec = register_int_counter_vec!(
@@ -190,5 +219,17 @@ lazy_static! {
     /// Counter of flush operations on puffin files.
     pub static ref INDEX_PUFFIN_FLUSH_OP_TOTAL: IntCounter = INDEX_IO_OP_TOTAL
         .with_label_values(&["flush", "puffin"]);
+    /// Counter of read operations on intermediate files.
+    pub static ref INDEX_INTERMEDIATE_READ_OP_TOTAL: IntCounter = INDEX_IO_OP_TOTAL
+        .with_label_values(&["read", "intermediate"]);
+    /// Counter of seek operations on intermediate files.
+    pub static ref INDEX_INTERMEDIATE_SEEK_OP_TOTAL: IntCounter = INDEX_IO_OP_TOTAL
+        .with_label_values(&["seek", "intermediate"]);
+    /// Counter of write operations on intermediate files.
+    pub static ref INDEX_INTERMEDIATE_WRITE_OP_TOTAL: IntCounter = INDEX_IO_OP_TOTAL
+        .with_label_values(&["write", "intermediate"]);
+    /// Counter of flush operations on intermediate files.
+    pub static ref INDEX_INTERMEDIATE_FLUSH_OP_TOTAL: IntCounter = INDEX_IO_OP_TOTAL
+        .with_label_values(&["flush", "intermediate"]);
     // ------- End of index metrics.
 }
diff --git a/src/mito2/src/row_converter.rs b/src/mito2/src/row_converter.rs
@@ -48,7 +48,7 @@ pub trait RowCodec {
     fn decode(&self, bytes: &[u8]) -> Result<Vec<Value>>;
 }
 
-#[derive(Debug)]
+#[derive(Debug, Clone, PartialEq, Eq)]
 pub struct SortField {
     data_type: ConcreteDataType,
 }

diff --git a/src/mito2/src/sst/index.rs b/src/mito2/src/sst/index.rs
@@ -16,6 +16,14 @@
 
 pub mod applier;
 mod codec;
+pub mod creator;
 mod store;
 
 const INDEX_BLOB_TYPE: &str = "greptime-inverted-index-v1";
+
+// TODO(zhongzc): how to determine this value?
+/// The minimum memory usage threshold for a column to qualify for external sorting during index creation.
+const MIN_MEMORY_USAGE_THRESHOLD: usize = 8192;
+
+/// The buffer size for the pipe used to send index data to the puffin blob.
+const PIPE_BUFFER_SIZE_FOR_SENDING_BLOB: usize = 8192;
diff --git a/src/mito2/src/sst/index/applier/builder.rs b/src/mito2/src/sst/index/applier/builder.rs
@@ -32,14 +32,13 @@ use index::inverted_index::search::predicate::Predicate;
 use object_store::ObjectStore;
 use snafu::{OptionExt, ResultExt};
 use store_api::metadata::RegionMetadata;
+use store_api::storage::ColumnId;
 
 use crate::error::{BuildIndexApplierSnafu, ColumnNotFoundSnafu, ConvertValueSnafu, Result};
 use crate::row_converter::SortField;
 use crate::sst::index::applier::SstIndexApplier;
 use crate::sst::index::codec::IndexValueCodec;
 
-type ColumnName = String;
-
 /// Constructs an [`SstIndexApplier`] which applies predicates to SST files during scan.
 pub struct SstIndexApplierBuilder<'a> {
     /// Directory of the region, required argument for constructing [`SstIndexApplier`].
@@ -52,7 +51,7 @@ pub struct SstIndexApplierBuilder<'a> {
     metadata: &'a RegionMetadata,
 
     /// Stores predicates during traversal on the Expr tree.
-    output: HashMap<ColumnName, Vec<Predicate>>,
+    output: HashMap<ColumnId, Vec<Predicate>>,
 }
 
 impl<'a> SstIndexApplierBuilder<'a> {
@@ -81,7 +80,11 @@ impl<'a> SstIndexApplierBuilder<'a> {
             return Ok(None);
         }
 
-        let predicates = self.output.into_iter().collect();
+        let predicates = self
+            .output
+            .into_iter()
+            .map(|(column_id, predicates)| (column_id.to_string(), predicates))
+            .collect();
         let applier = PredicatesIndexApplier::try_from(predicates);
         Ok(Some(SstIndexApplier::new(
             self.region_dir,
@@ -122,18 +125,16 @@ impl<'a> SstIndexApplierBuilder<'a> {
     }
 
     /// Helper function to add a predicate to the output.
-    fn add_predicate(&mut self, column_name: &str, predicate: Predicate) {
-        match self.output.get_mut(column_name) {
-            Some(predicates) => predicates.push(predicate),
-            None => {
-                self.output.insert(column_name.to_string(), vec![predicate]);
-            }
-        }
+    fn add_predicate(&mut self, column_id: ColumnId, predicate: Predicate) {
+        self.output.entry(column_id).or_default().push(predicate);
     }
 
-    /// Helper function to get the column type of a tag column.
+    /// Helper function to get the column id and the column type of a tag column.
     /// Returns `None` if the column is not a tag column.
-    fn tag_column_type(&self, column_name: &str) -> Result<Option<ConcreteDataType>> {
+    fn tag_column_id_and_type(
+        &self,
+        column_name: &str,
+    ) -> Result<Option<(ColumnId, ConcreteDataType)>> {
         let column = self
             .metadata
             .column_by_name(column_name)
@@ -142,7 +143,7 @@ impl<'a> SstIndexApplierBuilder<'a> {
             })?;
 
         Ok((column.semantic_type == SemanticType::Tag)
-            .then(|| column.column_schema.data_type.clone()))
+            .then(|| (column.column_id, column.column_schema.data_type.clone())))
     }
 
     /// Helper function to get a non-null literal.
@@ -303,15 +304,15 @@ mod tests {
         });
 
         builder.traverse_and_collect(&expr);
-        let predicates = builder.output.get("a").unwrap();
+        let predicates = builder.output.get(&1).unwrap();
         assert_eq!(predicates.len(), 1);
         assert_eq!(
             predicates[0],
             Predicate::RegexMatch(RegexMatchPredicate {
                 pattern: "bar".to_string()
             })
         );
-        let predicates = builder.output.get("b").unwrap();
+        let predicates = builder.output.get(&2).unwrap();
         assert_eq!(predicates.len(), 1);
         assert_eq!(
             predicates[0],

diff --git a/src/mito2/src/sst/index/applier/builder/between.rs b/src/mito2/src/sst/index/applier/builder/between.rs
@@ -28,7 +28,7 @@ impl<'a> SstIndexApplierBuilder<'a> {
         let Some(column_name) = Self::column_name(&between.expr) else {
             return Ok(());
         };
-        let Some(data_type) = self.tag_column_type(column_name)? else {
+        let Some((column_id, data_type)) = self.tag_column_id_and_type(column_name)? else {
             return Ok(());
         };
         let Some(low) = Self::nonnull_lit(&between.low) else {
@@ -51,7 +51,7 @@ impl<'a> SstIndexApplierBuilder<'a> {
             },
         });
 
-        self.add_predicate(column_name, predicate);
+        self.add_predicate(column_id, predicate);
         Ok(())
     }
 }
@@ -80,7 +80,7 @@ mod tests {
 
         builder.collect_between(&between).unwrap();
 
-        let predicates = builder.output.get("a").unwrap();
+        let predicates = builder.output.get(&1).unwrap();
         assert_eq!(predicates.len(), 1);
         assert_eq!(
             predicates[0],

diff --git a/src/mito2/src/sst/index/applier/builder/comparison.rs b/src/mito2/src/sst/index/applier/builder/comparison.rs
@@ -114,15 +114,15 @@ impl<'a> SstIndexApplierBuilder<'a> {
         let Some(lit) = Self::nonnull_lit(literal) else {
             return Ok(());
         };
-        let Some(data_type) = self.tag_column_type(column_name)? else {
+        let Some((column_id, data_type)) = self.tag_column_id_and_type(column_name)? else {
             return Ok(());
         };
 
         let predicate = Predicate::Range(RangePredicate {
             range: range_builder(Self::encode_lit(lit, data_type)?),
         });
 
-        self.add_predicate(column_name, predicate);
+        self.add_predicate(column_id, predicate);
         Ok(())
     }
 }
@@ -230,7 +230,7 @@ mod tests {
             builder.collect_comparison_expr(left, op, right).unwrap();
         }
 
-        let predicates = builder.output.get("a").unwrap();
+        let predicates = builder.output.get(&1).unwrap();
         assert_eq!(predicates.len(), cases.len());
         for ((_, expected), actual) in cases.into_iter().zip(predicates) {
             assert_eq!(