lotus/blockstore/splitstore
ZenGround0 544af751cb
Fix the overzealous fix ()
Co-authored-by: zenground0 <ZenGround0@users.noreply.github.com>
2023-02-28 09:22:18 -05:00
..
checkpoint_test.go test: use T.TempDir to create temporary test directory 2022-03-15 00:21:31 +08:00
checkpoint.go chore: remove redundant import prefixes 2022-06-15 12:06:22 +02:00
coldset_test.go test: use T.TempDir to create temporary test directory 2022-03-15 00:21:31 +08:00
coldset.go chore: remove redundant import prefixes 2022-06-15 12:06:22 +02:00
debug.go chore: all: bump go-libipfs to replace go-block-format 2023-01-26 17:03:18 +01:00
markset_badger.go Two fixes, better logging and comments () 2023-02-23 10:05:18 -05:00
markset_map.go chore: remove redundant import prefixes 2022-06-15 12:06:22 +02:00
markset_test.go fix make gen 2022-08-29 16:25:30 +02:00
markset.go chore: remove redundant import prefixes 2022-06-15 12:06:22 +02:00
README.md update README 2022-02-08 10:10:51 +02:00
splitstore_check.go feat:splitstore:single compaction that can handle prune aka two marksets one compaction () 2022-11-07 16:31:12 -05:00
splitstore_compact.go Fix the overzealous fix () 2023-02-28 09:22:18 -05:00
splitstore_expose.go chore: all: bump go-libipfs to replace go-block-format 2023-01-26 17:03:18 +01:00
splitstore_gc.go feat:chain:splitstore chain prune () 2022-08-05 16:34:16 -04:00
splitstore_prune.go Two fixes, better logging and comments () 2023-02-23 10:05:18 -05:00
splitstore_reify.go chore: all: bump go-libipfs to replace go-block-format 2023-01-26 17:03:18 +01:00
splitstore_test.go chore: all: bump go-libipfs to replace go-block-format 2023-01-26 17:03:18 +01:00
splitstore_util.go chore: remove redundant import prefixes 2022-06-15 12:06:22 +02:00
splitstore_warmup.go chore: all: bump go-libipfs to replace go-block-format 2023-01-26 17:03:18 +01:00
splitstore.go chore: all: bump go-libipfs to replace go-block-format 2023-01-26 17:03:18 +01:00
visitor.go chore: remove redundant import prefixes 2022-06-15 12:06:22 +02:00

SplitStore: An actively scalable blockstore for the Filecoin chain

The SplitStore was first introduced in lotus v1.5.1, as an experiment in reducing the performance impact of large blockstores.

With lotus v1.11.1, we introduce the next iteration in design and implementation, which we call SplitStore v1.

The new design (see #6474 evolves the splitstore to be a freestanding compacting blockstore that allows us to keep a small (60-100GB) working set in a hot blockstore and reliably archive out of scope objects in a coldstore. The coldstore can also be a discard store, whereby out of scope objects are discarded or a regular badger blockstore (the default), which can be periodically garbage collected according to configurable user retention policies.

To enable the splitstore, edit .lotus/config.toml and add the following:

[Chainstore]
  EnableSplitstore = true

If you intend to use the discard coldstore, your also need to add the following:

  [Chainstore.Splitstore]
    ColdStoreType = "discard"

In general you should not have to use the discard store, unless you are running a network assistive node (like a bootstrapper or booster) or have very constrained hardware with not enough disk space to maintain a coldstore, even with garbage collection. It is also appropriate for small nodes that are simply watching the chain.

Warning: Using the discard store for a general purpose node is discouraged, unless you really know what you are doing. Use it at your own risk.

Configuration Options

These are options in the [Chainstore.Splitstore] section of the configuration:

  • HotStoreType -- specifies the type of hotstore to use. The only currently supported option is "badger".
  • ColdStoreType -- specifies the type of coldstore to use. The default value is "universal", which will use the initial monolith blockstore as the coldstore. The other possible value is "discard", as outlined above, which is specialized for running without a coldstore. Note that the discard store wraps the initial monolith blockstore and discards writes; this is necessary to support syncing from a snapshot.
  • MarkSetType -- specifies the type of markset to use during compaction. The markset is the data structure used by compaction/gc to track live objects. The default value is "badger", which will use a disk backed markset using badger. If you have a lot of memory (48G or more) you can also use "map", which will use an in memory markset, speeding up compaction at the cost of higher memory usage. Note: If you are using a VPS with a network volume, you need to provision at least 3000 IOPs with the badger markset.
  • HotStoreMessageRetention -- specifies how many finalities, beyond the 4 finalities maintained by default, to maintain messages and message receipts in the hotstore. This is useful for assistive nodes that want to support syncing for other nodes beyond 4 finalities, while running with the discard coldstore option. It is also useful for miners who accept deals and need to lookback messages beyond the 4 finalities, which would otherwise hit the coldstore.
  • HotStoreFullGCFrequency -- specifies how frequenty to garbage collect the hotstore using full (moving) GC. The default value is 20, which uses full GC every 20 compactions (about once a week); set to 0 to disable full GC altogether. Rationale: badger supports online GC, and this is used by default. However it has proven to be ineffective in practice with the hotstore size slowly creeping up. In order to address this, we have added moving GC support in our badger wrapper, which can effectively reclaim all space. The downside is that it takes a bit longer to perform a moving GC and you also need enough space to house the new hotstore while the old one is still live.

Operation

When the splitstore is first enabled, the existing blockstore becomes the coldstore and a fresh hotstore is initialized.

The hotstore is warmed up on first startup so as to load all chain headers and state roots in the current head. This allows us to immediately gain the performance benefits of a smallerblockstore which can be substantial for full archival nodes.

All new writes are directed to the hotstore, while reads first hit the hotstore, with fallback to the coldstore.

Once 5 finalities have ellapsed, and every finality henceforth, the blockstore compacts. Compaction is the process of moving all unreachable objects within the last 4 finalities from the hotstore to the coldstore. If the system is configured with a discard coldstore, these objects are discarded. Note that chain headers, all the way to genesis, are considered reachable. Stateroots and messages are considered reachable only within the last 4 finalities, unless there is a live reference to them.

Compaction

Compaction works transactionally with the following algorithm:

  • We prepare a transaction, whereby all i/o referenced objects through the API are tracked.
  • We walk the chain and mark reachable objects, keeping 4 finalities of state roots and messages and all headers all the way to genesis.
  • Once the chain walk is complete, we begin full transaction protection with concurrent marking; we walk and mark all references created during the chain walk. On the same time, all I/O through the API concurrently marks objects as live references.
  • We collect cold objects by iterating through the hotstore and checking the mark set; if an object is not marked, then it is candidate for purge.
  • When running with a coldstore, we next copy all cold objects to the coldstore.
  • At this point we are ready to begin purging:
    • We sort cold objects heaviest first, so as to never delete the consituents of a DAG before the DAG itself (which would leave dangling references)
    • We delete in small batches taking a lock; each batch is checked again for marks, from the concurrent transactional mark, so as to never delete anything live
  • We then end the transaction and compact/gc the hotstore.

As of #8008 the compaction algorithm has been modified to eliminate sorting and maintain the cold object set on disk. This drastically reduces memory usage; in fact, when using badger as the markset compaction uses very little memory, and it should be now possible to run splitstore with 32GB of RAM or less without danger of running out of memory during compaction.

Garbage Collection

TBD -- see #6577

Utilities

lotus-shed has a splitstore command which provides some utilities:

  • rollback -- rolls back a splitstore installation. This command copies the hotstore on top of the coldstore, and then deletes the splitstore directory and associated metadata keys. It can also optionally compact/gc the coldstore after the copy (with the --gc-coldstore flag) and automatically rewrite the lotus config to disable splitstore (with the --rewrite-config flag). Note: the node must be stopped before running this command.
  • clear -- clears a splitstore installation for restart from snapshot.
  • check -- asynchronously runs a basic healthcheck on the splitstore. The results are appended to <lotus-repo>/datastore/splitstore/check.txt.
  • info -- prints some basic information about the splitstore.