Following notation is used: words written in italics and wihout spacing mean class name without package name or method name, for example, GridCacheMapEntry

Table of Contents:

Persistence and Crash Recovery

Crash Recovery can be

• Local (most DB are able to do this)
• and distributed (whole cluster state is restored).

Local Crash Recovery

Ignite Durable Memory is basis for all data structures. There is no cache state saved on heap now. 

To save cache state to disk we can dump all its pages to disk. First prototypes used this simple approach: stop all updates and save all pages.

WAL

We can’t control moment when node crashes. Also we can’t save all memory pages to disk each time - it is too slow. Updates should be incremental. 

Let's suppose we have saved tree leafs, but didn’t save tree root (during pages allocation they may be reordered because allocation is multithread). In this case all updates will be lost.

Technique to solve this named write ahead logging: Before doing actual update, we append planned change information into cyclic file named WAL log (operation name - WAL append/WAL log).

After crash we can read and replay WAL using already saved page set. We can restore to state, which was last committed state of crashed process. Restore bases on pages set + WAL. See also WAL structure section below

Practically we can’t replay WAL from the beginning of times, Volume(HDD)<Volume(full WAL), and we need procedure to throw out oldest part of changes in WAL.

This procedure is named checkpointing

Checkpointing

Can be of two types

• Sharp Checkpointing - if checkpoint is completed all data structures on disk are consistent, data is consistent in terms of references and transactions.
• Fuzzy Checkpointing - means state on disk may require recovery itself

Implemented - Sharp Checkpoint; F.C. - todo

To achieve consistency Checkpoint read-write lock is used (see GridCacheDatabaseSharedManager#checkpointLock)

• Cache Updates - holds read lock
• Checkpointer - holds write lock for short time. Holding write lock means all state is consistent, updates are not possible. Usage of CP Lock allows to do sharp checkpoint

Under CP write lock held we do the following:

1. 1. 1. WAL marker is added: CP (begin) record is added - CheckpointRecord - marks consistent state in WAL
      2. Collect pages were changed since last checkpoint

And then CP write lock is released, updates and transactions can run.

Dirty pages is set, when page from non-dirty becomes dirty, it is added to this set.

Collection of pages (GridCacheDatabaseSharedManager.Checkpoint#cpPages) allows us to collect and then write pages which were changed since last checkpoint.

Checkpoint Pool

In parallel with process of writing pages to disk, some thread may want to update data in the page being written.

For such case Checkpoint pool is used for pages being updated in parallel with write. This pool has limitation.

Copy on write technique is used. If there is modification in page which is under CP now we will create temporary copy of page.

Triggers

• Percent of dirty pages is trigger for checkpointing (e.g. 75%).
• Timeout is also trigger, do checkpoint every N seconds

WAL structure

WAL file segments and rotation structure

See also WAL history size section below

WAL records for recovery

Crash recovery involves following records writtent in WAL, it may be of 2 main types

1. Logical record
   1. Operation description - which operation we want to do. Contains operation type (put, remove) and (Key, Value, Version)  - DataRecord
   2. Transactional record - begin, prepare, commit, and rollback tx records - (TxRecord)
2. Physical records
   1. Delta record - describes memory region change, page change. Subclass of PageDeltaRecord. Contains bytes changed in the page. e.g bytes 5-10 were changed to [...,]. Relatively small records for B+tree records
   2. Full page snapshot - written for first page change after CP, when page state changes from clean->dirty state (PageSnapshot)

For particular update we write records in follwowing order

1. logical record with change planned - DataRecord with several DataEntry (ies)
2. page record:
   
   1. option: then for page changed by this update we write initially clean - PageSnapshot,
   2. option: for already modified - PageDeltaRecord

Possible future optimisation - refer data modified from PageDeltaRecord to logical record. Will allow to not store byte updates twice. We have file wal pointer, pointer to record from the beginning of time. This refreence may be used.

Local Recovery Process

Let’s assume node start process is running with existent files.

1. We need to check if page store is consistent.
2. Or we need to find out if crash was while Checkpoint (CP) was running
   

Ignite manages 2 types of CP markers on disk (standalone files, includes timestamp and WAL pointer):

• CP begin
• CP end

If we observe only CP begin and there is no CP end marker that means CP not finished; we have not consistent page store.

Let’s suppose we discover markers for CP1 and 2 start and CP 1 end.

For completed checkpoint we apply only physical records, for incomplete  - only logical (as physical may be corrupted).

Page Snapshot records required to avoid double apply of data from delta records.

When replay is finished CP2 marker will be added.

If transaction begin record has no corresponding end, tx change is not applied.

Summary, limitations and performance

Persistence files

There are next file types for DB

• WAL segments - constant size file (WAL work directory 0...9.wal;, WAL archive 0.wal…)
• CP markers (UUID-Begin.bin, UUID-End.bin)
• Page store (Now implemented as file per partition: cache_name\part1,2,3.bin)

Consistent state comes only from pair of WAL and page store.

Limitations

Because CP are consistent we can’t start next CP until previous is not completed.

There is possible next situation:

• updates coming fast from worker threads
• CP pool (for copy on writes) may become full with new changes originated

For that case we will block new updates and wait running for CP to finish.

To avoid such scenario:

• increase frequency of checkpoints (to minimize amount of data to be saved in each CP)
• increase CP buffer size

WAL and page store may be saved to different devices to avoid its mutual influence.

Case if same records are updated many times may generate load to WAL and no significant load to page store.

To provide recovery guarantees each write (log()) to WAL should:

• call write() itself.
• but also require fsync (force buffers to be flushed by OS to the real device).

fsync is expensive operation. There is optimisation for case updates coming faster than disk write, fsyncDelayNanos (1ns-1ms, 1ns by default) delay is used. This delay is used to park threads to accumulate more than one fsync requests.

Future optimisation: standalone thread will be responsible to write data to disk. Worker threads will do all preparation and transfer buffer to write.

See also WAL history size section below.

WAL mode

There several levels of guarantees (WALMode)

But there is several nodes containing same data and there is possible to restore data from other nodes.

Distributed Recovery

Partition update counter. This mechanism was already used in continuous queries.

• Partition update counter is associated with partition
• Each update causes increment of partition update counter.
  

Each update (counter) is replicated to backup. If counter equal on primary and backup means replication is finished.

Partition update counter is saved with update recods in WAL.

Node Join (with data from persitence)

Consider partition on joining node was is owning state, update counter = 50. Existing nodes has update counter = 150

Node join causes partition map exchange, update counter is sent with other partition data. (Joining node will have new ID and from the point of view of dicsovery this node is a new node.)

Coordinator observes older partition state and forces partition to moving state. Moving force is required to setup uploading newer data.

Rebalance of fresh data to joined node now may be run in 2 modes:

• There is WAL on primary node. WAL includes checkpoint marker with partition update cntr = 45. 
  • We can send only WAL logical update records to backup
• If counter in WAL is too big, e.g. 200, we don’t have delta (can't sent WAL recods) 
  • joined node will have to clear partition data. 
  • Partition state is set to renting state
  • When clean up finished partition goes to moving state.
  • We can’t use delta updates because there is possible problem with keys deleted early. Can get stale key if we send only delta of changes.

Possible future optimisation: for full update we may send page store file over network.

WAL history size

In corner case we need to store WAL only for 1 checkpoint in past for successful recovery (PersistentStoreConfiguration#walHistSize )

We can’t delete WAL segments considering only history size in bytes or segments. It is possible to replay WAL only starting from checkpoint marker.

Assuming that checkpoints are triggered mostly by timeout we can estimate possible downtime after which node may be rebalanced using delta logical WAL records.

By default WAL history size is 20 to increase probability that rebalancing can be done using logical deltas from WAL.