...

• A complex index reduild procedure that requires the development of additional crash recovery guarantees. It will start immediately when the partition file is fully received from the supplier node. If the node crashes in the middle of the rebuilding index process it will have an inconsistent index state at the further node startup. To avoid this a new index-undo WAL record must be logged within rebuilding and used on node start to remove previously added index records.

...

Historical rebalance

...

After partition is received the historical rebalance must be initiated to load other cache updates.

...

Catch-up temporary WAL

The swapped temporary storage will log all the cache updates to the temporary WAL storage (per each partition) for further applying them to the corresponding partition file.

Preload entries from loaded partition file

The demander node will use a preloaded patition file as a new source of cache data entries to load.

Disadvantages:

• The approach will require a new temporary FilePageStore to be initialized. It must be created as a part of the temporary cache group or in the separate temporary data region to provide reusing machinery of iteration over the full partition file.

Proposed Changes (Hot swap approach with historical rebalance)

Process Overview

In the process of balancing data:

• demaner (receiver of partition files)
• supplier (sender of partition files).

The whole process is described in terms of rebalance single cache group and partition files would be rebalanced one-by-one:

1. The demander node prepares the set of IgniteDhtDemandedPartitionsMap#full cache partitions to fetch;
2. The demander node checks compatibility version (for example, 2.8) and starts recording all incoming cache updates to the new special storage – the temporary WAL storage;
3. The demander node sends the GridDhtPartitionDemandMessage to the supplier node as usual;
4. The supplier node receives GridDhtPartitionDemandMessage and starts the new checkpoint process and fixes cache partition file sizes;
5. The supplier node creates an empty temporary file with .delta (e.g. part-0.bin.delta file) postfix for each cache partition file (in the same cache working directory or another configured);
6. The supplier node starts tracking each pageId write attempt to these partition files
   
   1. When the write attempt happens, the thread that caused it reads the original copy of this page from the partition and flushes it to the corresponding .delta file;
   2. After it the thread writes the changed page data to the partition file;
7. The supplier waits the checkpoint process ends;
8. On the supplier for each cache partition file
   1. The process opens the partition file in read-only mode;
   2. Starts sending partition file (as it is with any concurrent writes) by chunks of predefined constant size (multiple of PageMemory size);
      
   3. After the partition file sent it starts sending corresponding .delta file;
9. The demander node starts to listen to new type of incoming connections (a socket channel created event) from the supplier node;
10. When the appropriate connection established the demander node for each cache partition file
    1. Receives file metadata information (corresponding cache group identifier, cache partition file name, file size)
    2. Writes data from the socket to the particular cache partition file from the beginning of the file
    3. After the original cache partition file received the node starts receiving corresponding .delta file
    4. The node reads data from the socket by chunks of PageMemory size and applies each received pageId to the partition file
11. When the demander node receives the whole cache partition file
    1. The node begins the rebuild secondary indexes procedure over received partition file
    2. After it the thread begins to apply for data entries from the beginning of WAL-temporary storage;
    3. All async operations corresponding to this partition file still write to the end of temporary WAL;
    4. At the moment of WAL-temporary storage is ready to be empty
       1. Start the first checkpoint;
       2. Wait for the first checkpoint ends and own the cache partition;
       3. All operations now are switched to the partition file instead of writing to the temporary WAL;
       4. Schedule the temporary WAL storage deletion;
12. The supplier node deletes all temporary files;

Components

In terms of a high-level abstraction, Apache Ignite must support the features described below.

File transfer between nodes

The node partition preloader machinery download cache partition files from cluster nodes which owns desired partitions (the zero copy algorithm [1] assume to be used by default). To achieve this, the file transmission process must be implemented at Apache Ignite over Communication SPI.

CommunicationSpi

IThe Comminication SPI must support to:

• opening channel connections to a remote node to an arbitrary topic (GridTopic is used) with initial meta information;
• listening incoming channel connections and handling them by registered handlers;
• an arbitrary set of channel parameters on connection handshake (some initial Message assumed to be used);

API

public interface CommunicationListenerEx<T extends Serializable> extends EventListener {
    /**
     * @param nodeId Remote node id.
     * @param initMsg Init channel message.
     * @param channel Locally created channel endpoint.
     */
    public void onChannelOpened(UUID nodeId, Message initMsg, Channel channel);
}

GridIoManager

While the Demander is being receive partition files it must save sequentially all cache entries corresponding to the MOVING partition into a new temporary storage. These entries will be applied later one by one on the newly received cache partition file. All asynchronous operations will be enrolled to the end of temporary storage during storage reads until it becomes fully read. The file-based FIFO approach assumes to be used by this process.

The temporary storage is chosen to be WAL-based. The storage must support to:

• Unlimited number of WAL-files to store temporary data records;
• Iterating over stored data records during an asynchronous writer thread insert new records;
• WAL-per-partiton approach is need to be used;
• Write operations to storage must have higher priority over read operations;

Expected problems to be solved

• We must stop updating indexes on demander when the data is ready to be transferred from the supplier node. All async cache updates on demander must not cause the index update;
• The previous partition metadata page and all stored meta information must be destroyed in PageMemory and restored from the new partition file;

Preload entries from loaded partition file

The demander node will use a preloaded patition file as a new source of cache data entries to load.

Disadvantages:

• The approach will require a new temporary FilePageStore to be initialized. It must be created as a part of the temporary cache group or in the separate temporary data region to provide reusing machinery of iteration over the full partition file.

Proposed Changes (Hot swap with historical rebalance)

Process Overview

In the process of balancing data:

• Demander (receiver of partition files)
• Supplier (sender of partition files)

The whole process is described in terms of rebalancing a single partition file of a cache group. All the other partitions would be rebalanced one-by-one.

1. NODE_JOIN event occurrs and the blocking PME starts;
   1. The Demander decides which partitions must be loaded. All the desired partitions have MOVING state;
   2. The Demander initiates a new checkpoint process;
      1. Under the checkpoint write-lock it swaps cache data storage with the temporary one for each partition of the given set;
      2. The temporary cache data storage tracks partition counter number as ususal (on each cache operations);
      3. Wait for the checkpoint begin future ends;
2. The Demander sends a request to the Supplier with the previously prepares set of cache groups and partition files;
3. The Supplier receives a request and starts a new local checkpoint process;
   1. Creates a temporary file with .delta postfix (for each partition file e.g. part-0.bin.delta);
   2. Under checkpoint write lock fixes the partition expected file size (at the moment of the checkpoint end);
   3. Wait for the checkpoint begin future ends;
   4. Starts the copy process of the partition file to the Demander;
      1. Opens the partition file in read-only mode;
      2. Starts sending partition file (with any concurrent writes) by chunks of predefined size;
   5. Asynchronouosly writes each page to the partition file and the same page to the corresponding file with .delta postfix;
   6. When the partition file sent it starts sending corresponding .delta file;
4. The Demander listens new file sending attempts from the Supplier;
5. The Demander receives partition file (for each partition file one by one);
6. The Demander reads corresponding partition .delta file by chunks and applies them on the received partiton file;
7. When the Demander receives the whole cache partition file;
   1. Swap the temporary cache data storage with the original one on the next checkpoint (under write lock); 
   2. When the partition has been swapped it starts the rebuild indexes procedure over given partition files;
   3. Starts historical rebalance for the given partition file;
8. The Supplier deletes all temporary files;

Components

In terms of a high-level abstraction, Apache Ignite must support the features described below.

File transfer between nodes

The node partition preloader machinery download cache partition files from cluster nodes which owns desired partitions (the zero copy algorithm [1] assume to be used by default). To achieve this, the file transmission process must be implemented at Apache Ignite over Communication SPI.

CommunicationSpi

IThe Comminication SPI must support to:

• opening channel connections to a remote node to an arbitrary topic (GridTopic is used) with initial meta information;
• listening incoming channel connections and handling them by registered handlers;
• an arbitrary set of channel parameters on connection handshake (some initial Message assumed to be used);

API

public interface CommunicationListenerEx<T extends Serializable> extends EventListener {
    /**
     * @param nodeId Remote node id.
     * @param initMsg Init channel message.
     * @param channel Locally created channel endpoint.
     */
    public void onChannelOpened(UUID nodeId, Message initMsg, Channel channel);
}

GridIoManager

IO manager must support to:IO manager must support to:

• different approaches of incoming data handling: CHUNK (read channel into ByteBuffer), FILE (zero-copy approach)
• send and receive data by chunks of predefined size with storing intermediate results;
• reestablishing connection between nodes if an error occurs and continue file sending\receiving;
• limiting connection bandwidth at runtime;

...

public class TransmissionSender implements Closeable {
    /**
     * @param file Source file to send to remote.
     * @param params Additional transfer file description keys.
     * @param plc The policy of handling data on remote.
     * @throws IgniteCheckedException If fails.
     */
    public void send(
        File file,
        Map<String, Serializable> params,
        TransmissionPolicy plc
    ) throws IgniteCheckedException, InterruptedException, IOException {
        send(file, 0, file.length(), params, plc);
    }

    /**
     * @param file Source file to send to remote.
     * @param plc The policy of handling data on remote.
     * @throws IgniteCheckedException If fails.
     */
    public void send(
        File file,
        TransmissionPolicy plc
    ) throws IgniteCheckedException, InterruptedException, IOException {
        send(file, 0, file.length(), new HashMap<>(), plc);
    }

    /**
     * @param file Source file to send to remote.
     * @param offset Position to start trasfer at.
     * @param cnt Number of bytes to transfer.
     * @param params Additional transfer file description keys.
     * @param plc The policy of handling data on remote.
     * @throws IgniteCheckedException If fails.
     */
    public void send(
        File file,
        long offset,
        long cnt,
        Map<String, Serializable> params,
        TransmissionPolicy plc
    ) throws IgniteCheckedException, InterruptedException, IOException {
		// Impl.
    }
}

Copy partition on the fly

Checkpointer

When the supplier node receives the cache partition file demand request it will send the file over the CommunicationSpi. The cache partition file can be concurrently updated by checkpoint thread during its transmission. To guarantee the file consistency Сheckpointer must use Copy-on-Write [3] tehnique and save a copy of updated chunk into the temporary file.

Apply partition on the fly

Catch-up temporary WAL

While the demander node is in the partition file transmission state it must save sequentially all cache entries corresponding to the MOVING partition into a new temporary storage. These entries will be applied later one by one on the newly received cache partition file. All asynchronous operations will be enrolled to the end of temporary storage during storage reads until it becomes fully read. The file-based FIFO approach assumes to be used by this process.

The temporary storage is chosen to be WAL-based. The storage must support to:

• Unlimited number of WAL-files to store temporary data records;
• Iterating over stored data records during an asynchronous writer thread insert new records;
• WAL-per-partiton approach is need to be used;
• Write operations to storage must have higher priority over read operations;

Expected problems to be solved

...

   }
}

Copy partition on the fly

Checkpointer

When the supplier node receives the cache partition file demand request it will send the file over the CommunicationSpi. The cache partition file can be concurrently updated by checkpoint thread during its transmission. To guarantee the file consistency Сheckpointer must use Copy-on-Write [3] tehnique and save a copy of updated chunk into the temporary file.

...

Rebuild indexes

The node is ready to become partition owner when partition data is rebalanced and cache indexes are ready. For the message-based cluster rebalancing approach indexes are rebuilding simultaneously with cache data loading. For the file-based rebalancing approach, the index rebuild procedure must be finished before the partition state is set to the OWNING state. 

Failover and Recovery

procedure must be finished before the partition state is set to the OWNING state. 

Failover and Recovery

Ignite doesn't provide any recovery guarantees for the partitions with the MOVING state. The cache partitions will be fully loaded when the next rebalance procedure occurs.

FAIL\LEFT during rebalancing

The node which is beeing rebalancing left the cluster. For such nodes WAL is always disabled (all partitions have MOVING state due to this node is new for the cluster and has no cache data). 
Since WAL is disabled we can guarantee that all operations with loaded partition files are safe to be done (renaming partition files, applying async updates) due to a cache directory will be fully dropped on recoveryApache Ignite doesn't provide any recovery guarantees for the partitions with the MOVING state. The cache partitions will be fully loaded when the next rebalance procedure occurs.

Topology change

Each topology change event JOIN/LEFT/FAILED may or may not change cache affinity assignments of currently rebalacning caches. If assignments is not changed and the node is still needs partitions being rebalanced we can continue the current rebalance process (see for details IGNITE-7165).

...

To provide basic recovery guarantees we must to: 

...

• Wait for the first checkpoint ends and set OWNING status to partition;

Recovery from different stages:

• The Supplier crashes when sending partition;
• The Demander crashes when receiving partition;
• The Demander crashes when applying temp WAL;

Phase-2

The SSL must be disabled to take an advantage of Java NIO zero-copy file transmission using FileChannel#transferTo method. If we need to use SSL the file must be splitted on chunks the same way to send them over the socket channel with ByteBuffer. As the SSL engine generally needs a direct ByteBuffer to do encryption we can't avoid copying buffer payload from the kernel level to the application level. 

...