[storage] Enable the strictest sync mode for RocksDB's WAL

This commit enables sync mode = true for RocksDB WAL, when committing a WriteBatch.
This option isn't enabled by default.
This would increase the durability level to survive a whole machine
crush, and with the expense of (potentially) higher commit latency.
Keep in mind that the WAL writes are temporarily enabled for the main
store, up to the point that our distributed log will be responsible for
durability.

crates/storage-rocksdb/src/lib.rs

@@ -33,7 +33,6 @@ use futures::{ready, FutureExt, Stream};
 use futures_util::future::ok;
 use futures_util::StreamExt;
 use restate_storage_api::{GetFuture, GetStream, PutFuture, Storage, StorageError, Transaction};
-use rocksdb::BlockBasedOptions;
 use rocksdb::Cache;
 use rocksdb::ColumnFamily;
 use rocksdb::DBCompressionType;
@@ -44,6 +43,7 @@ use rocksdb::PrefixRange;
 use rocksdb::ReadOptions;
 use rocksdb::SingleThreaded;
 use rocksdb::WriteBatch;
+use rocksdb::{BlockBasedOptions, WriteOptions};
 use std::pin::Pin;
 use std::sync::Arc;
 use std::task::{Context, Poll};
@@ -599,7 +599,9 @@ impl<'a> Transaction for RocksDBTransaction<'a> {
             // for now we always use the WAL when committing the write batch,
             // but once we will have the Raft log persisted else where, this becomes
             // only needed during a checkpoint operation.
-            db.write(write_batch)
+            let mut opts = WriteOptions::default();
+            opts.set_sync(true);
+            db.write_opt(write_batch, &opts)
                 .map_err(|error| StorageError::Generic(error.into()))
         };
         tokio::task::spawn_blocking(f)