Reducing the number of call to fsync on the directory.

This work by introducing a new API method in the Directory
trait.
The user needs to explicitely call this method.
(In particular, once before a commmit)

Closes #1225

CHANGELOG.md

@@ -2,6 +2,8 @@ Tantivy 0.17
 ================================
 - Change to non-strict schema. Ignore fields in data which are not defined in schema. Previously this returned an error. #1211
 - Facets are necessarily indexed. Existing index with indexed facets should work out of the box. Index without facets that are marked with index: false should be broken (but they were already broken in a sense). (@fulmicoton) #1195 .
+- Bugfix that could in theory impact durability in theory on some filesystems [#1224](https://github.com/quickwit-inc/tantivy/issues/1224)
+- Reduce the number of fsync calls [#1225](https://github.com/quickwit-inc/tantivy/issues/1225)
 
 Tantivy 0.16.2
 ================================

src/directory/directory.rs

@@ -142,10 +142,16 @@ pub trait Directory: DirectoryClone + fmt::Debug + Send + Sync + 'static {
     /// Opens a writer for the *virtual file* associated with
     /// a Path.
     ///
-    /// Right after this call, the file should be created
-    /// and any subsequent call to `open_read` for the
+    /// Right after this call, for the span of the execution of the program
+    /// the file should be created and any subsequent call to `open_read` for the
     /// same path should return a `FileSlice`.
     ///
+    /// However, depending on the directory implementation,
+    /// it might be required to call `sync_directory` to ensure
+    /// that the file is durably created.
+    /// (The semantics here are the same when dealing with
+    /// a posix filesystem.)
+    ///
     /// Write operations may be aggressively buffered.
     /// The client of this trait is responsible for calling flush
     /// to ensure that subsequent `read` operations
@@ -176,6 +182,12 @@ pub trait Directory: DirectoryClone + fmt::Debug + Send + Sync + 'static {
     /// The file may or may not previously exist.
     fn atomic_write(&self, path: &Path, data: &[u8]) -> io::Result<()>;
 
+    /// Sync the directory.
+    ///
+    /// This call is required to ensure that newly created files are
+    /// effectively stored durably.
+    fn sync_directory(&self) -> io::Result<()>;
+
     /// Acquire a lock in the given directory.
     ///
     /// The method is blocking or not depending on the `Lock` object.

src/directory/managed_directory.rs

@@ -192,6 +192,7 @@ impl ManagedDirectory {
             for delete_file in &deleted_files {
                 managed_paths_write.remove(delete_file);
             }
+            self.directory.sync_directory()?;
             save_managed_paths(self.directory.as_mut(), &meta_informations_wlock)?;
         }
 
@@ -222,9 +223,21 @@ impl ManagedDirectory {
             .write()
             .expect("Managed file lock poisoned");
         let has_changed = meta_wlock.managed_paths.insert(filepath.to_owned());
-        if has_changed {
-            save_managed_paths(self.directory.as_ref(), &meta_wlock)?;
+        if !has_changed {
+            return Ok(());
+        }
+        save_managed_paths(self.directory.as_ref(), &meta_wlock)?;
+        if meta_wlock.managed_paths.len() >= 1 {
+            // This is not the first file we add.
+            // `.managed.json` has been already properly created and we do not need to
+            // sync its parent directory.
+            //
+            // Note we do not try to create the managed.json file, upon the
+            // creation of the ManagedDirectory because it would prevent
+            // the usage of a ReadOnly directory.
+            return Ok(());
         }
+        self.directory.sync_directory()?;
         Ok(())
     }
 
@@ -310,6 +323,11 @@ impl Directory for ManagedDirectory {
     fn watch(&self, watch_callback: WatchCallback) -> crate::Result<WatchHandle> {
         self.directory.watch(watch_callback)
     }
+
+    fn sync_directory(&self) -> io::Result<()> {
+        self.directory.sync_directory()?;
+        Ok(())
+    }
 }
 
 impl Clone for ManagedDirectory {

src/directory/mmap_directory.rs

@@ -211,33 +211,6 @@ impl MmapDirectory {
         self.inner.root_path.join(relative_path)
     }
 
-    /// Sync the root directory.
-    /// In certain FS, this is required to persistently create
-    /// a file.
-    fn sync_directory(&self) -> Result<(), io::Error> {
-        let mut open_opts = OpenOptions::new();
-
-        // Linux needs read to be set, otherwise returns EINVAL
-        // write must not be set, or it fails with EISDIR
-        open_opts.read(true);
-
-        // On Windows, opening a directory requires FILE_FLAG_BACKUP_SEMANTICS
-        // and calling sync_all() only works if write access is requested.
-        #[cfg(windows)]
-        {
-            use std::os::windows::fs::OpenOptionsExt;
-            use winapi::um::winbase;
-
-            open_opts
-                .write(true)
-                .custom_flags(winbase::FILE_FLAG_BACKUP_SEMANTICS);
-        }
-
-        let fd = open_opts.open(&self.inner.root_path)?;
-        fd.sync_data()?;
-        Ok(())
-    }
-
     /// Returns some statistical information
     /// about the Mmap cache.
     ///
@@ -367,22 +340,17 @@ impl Directory for MmapDirectory {
     /// removed before the file is deleted.
     fn delete(&self, path: &Path) -> result::Result<(), DeleteError> {
         let full_path = self.resolve_path(path);
-        match fs::remove_file(&full_path) {
-            Ok(_) => self.sync_directory().map_err(|e| DeleteError::IoError {
-                io_error: e,
-                filepath: path.to_path_buf(),
-            }),
-            Err(e) => {
-                if e.kind() == io::ErrorKind::NotFound {
-                    Err(DeleteError::FileDoesNotExist(path.to_owned()))
-                } else {
-                    Err(DeleteError::IoError {
-                        io_error: e,
-                        filepath: path.to_path_buf(),
-                    })
+        fs::remove_file(&full_path).map_err(|e| {
+            if e.kind() == io::ErrorKind::NotFound {
+                DeleteError::FileDoesNotExist(path.to_owned())
+            } else {
+                DeleteError::IoError {
+                    io_error: e,
+                    filepath: path.to_path_buf(),
                 }
             }
-        }
+        })?;
+        Ok(())
     }
 
     fn exists(&self, path: &Path) -> Result<bool, OpenReadError> {
@@ -411,10 +379,13 @@ impl Directory for MmapDirectory {
         file.flush()
             .map_err(|io_error| OpenWriteError::wrap_io_error(io_error, path.to_path_buf()))?;
 
-        // Apparetntly, on some filesystem syncing the parent
-        // directory is required.
-        self.sync_directory()
-            .map_err(|io_err| OpenWriteError::wrap_io_error(io_err, path.to_path_buf()))?;
+        // Note we actually do not sync the parent directory here.
+        //
+        // A newly created file, may, in some case, be created and even flushed to disk.
+        // and then lost...
+        //
+        // The file will only be durably written after we terminate AND
+        // sync_directory() is called.
 
         let writer = SafeFileWriter::new(file);
         Ok(BufWriter::new(Box::new(writer)))
@@ -444,7 +415,7 @@ impl Directory for MmapDirectory {
         debug!("Atomic Write {:?}", path);
         let full_path = self.resolve_path(path);
         atomic_write(&full_path, content)?;
-        self.sync_directory()
+        Ok(())
     }
 
     fn acquire_lock(&self, lock: &Lock) -> Result<DirectoryLock, LockError> {
@@ -470,6 +441,30 @@ impl Directory for MmapDirectory {
     fn watch(&self, watch_callback: WatchCallback) -> crate::Result<WatchHandle> {
         Ok(self.inner.watch(watch_callback))
     }
+
+    fn sync_directory(&self) -> Result<(), io::Error> {
+        let mut open_opts = OpenOptions::new();
+
+        // Linux needs read to be set, otherwise returns EINVAL
+        // write must not be set, or it fails with EISDIR
+        open_opts.read(true);
+
+        // On Windows, opening a directory requires FILE_FLAG_BACKUP_SEMANTICS
+        // and calling sync_all() only works if write access is requested.
+        #[cfg(windows)]
+        {
+            use std::os::windows::fs::OpenOptionsExt;
+            use winapi::um::winbase;
+
+            open_opts
+                .write(true)
+                .custom_flags(winbase::FILE_FLAG_BACKUP_SEMANTICS);
+        }
+
+        let fd = open_opts.open(&self.inner.root_path)?;
+        fd.sync_data()?;
+        Ok(())
+    }
 }
 
 #[cfg(test)]

src/directory/ram_directory.rs

@@ -225,6 +225,10 @@ impl Directory for RamDirectory {
     fn watch(&self, watch_callback: WatchCallback) -> crate::Result<WatchHandle> {
         Ok(self.fs.write().unwrap().watch(watch_callback))
     }
+
+    fn sync_directory(&self) -> io::Result<()> {
+        Ok(())
+    }
 }
 
 #[cfg(test)]

src/indexer/segment_updater.rs

@@ -1,4 +1,5 @@
 use super::segment_manager::SegmentManager;
+use crate::TantivyError;
 use crate::core::Index;
 use crate::core::IndexMeta;
 use crate::core::IndexSettings;
@@ -20,7 +21,6 @@ use crate::indexer::{DefaultMergePolicy, MergePolicy};
 use crate::indexer::{MergeCandidate, MergeOperation};
 use crate::schema::Schema;
 use crate::Opstamp;
-use crate::TantivyError;
 use fail::fail_point;
 use futures::channel::oneshot;
 use futures::executor::{ThreadPool, ThreadPoolBuilder};
@@ -61,7 +61,9 @@ pub fn save_new_metas(
             payload: None,
         },
         directory,
-    )
+    )?;
+    directory.sync_directory()?;
+    Ok(())
 }
 
 /// Save the index meta file.
@@ -82,6 +84,7 @@ fn save_metas(metas: &IndexMeta, directory: &dyn Directory) -> crate::Result<()>
         io::ErrorKind::Other,
         msg.unwrap_or_else(|| "Undefined".to_string())
     ))));
+    directory.sync_directory()?;
     directory.atomic_write(&META_FILEPATH, &buffer[..])?;
     debug!("Saved metas {:?}", serde_json::to_string_pretty(&metas));
     Ok(())

src/store/compression_lz4_block.rs

@@ -1,14 +1,15 @@
 use std::io::{self};
 
 use core::convert::TryInto;
-use std::mem;
 use lz4_flex::{compress_into, decompress_into};
+use std::mem;
 
 #[inline]
 #[allow(clippy::uninit_vec)]
 pub fn compress(uncompressed: &[u8], compressed: &mut Vec<u8>) -> io::Result<()> {
     compressed.clear();
-    let maximum_ouput_size = mem::size_of::<u32>() + lz4_flex::block::get_maximum_output_size(uncompressed.len());
+    let maximum_ouput_size =
+        mem::size_of::<u32>() + lz4_flex::block::get_maximum_output_size(uncompressed.len());
     compressed.reserve(maximum_ouput_size);
     unsafe {
         compressed.set_len(maximum_ouput_size);