This commit
- adds a configuration in which no step holds, replication cursors, etc. are created
- removes the send.step_holds.disable_incremental setting
- creates a new config option `replication` for active-side jobs
- adds the replication.protection.{initial,incremental} settings, each
of which can have values
- `guarantee_resumability`
- `guarantee_incremental`
- `guarantee_nothing`
(refer to docs/configuration/replication.rst for semantics)
The `replication` config from an active side is sent to both endpoint.Sender and endpoint.Receiver
for each replication step. Sender and Receiver then act accordingly.
For `guarantee_incremental`, we add the new `tentative-replication-cursor` abstraction.
The necessity for that abstraction is outlined in https://github.com/zrepl/zrepl/issues/340.
fixes https://github.com/zrepl/zrepl/issues/340
This is a stop-gap solution until we re-write the pruner to support
rules for removing step holds.
Note that disabling step holds for incremental sends does not affect
zrepl's guarantee that incremental replication is always possible:
Suppose you yank the external drive during an incremental @from -> @to step:
* restarting that step or future incrementals @from -> @to_later` will be possible
because the replication cursor bookmark points to @from until the step is complete
* resuming @from -> @to will work as long as the pruner on your internal pool doesn't come around to destroy @to.
* in that case, the replication algorithm should determine that the resumable state
on the receiving side isuseless because @to no longer exists on the sending side,
and consequently clear it, and restart an incremental step @from -> @to_later
refs #288
- drop HintMostRecentCommonAncestor rpc call
- it is wrong to put faith into the active side of the replication to always make that call
(we might not trust it, ref pull setup)
- clean up step holds + step bookmarks + replication cursor bookmarks on
send RPC instead
- this makes it symmetric with Receive RPC
- use a cache (endpoint.sendAbstractionsCache) to avoid the cost of
listing the on-disk endpoint abstractions state on every step
The "create" methods for endpoint abstractions (CreateReplicationCursor, HoldStep) are now fully
idempotent and return an Abstraction.
Notes about endpoint.sendAbstractionsCache:
- fills lazily from disk state on first `Get` operation
- fill from disk is generally only attempted once
- unless the `ListAbstractions` fails, in which case the fill from
disk is retried on next `Get` (the current `Get` will observe a
subset of the actual on-disk abstractions)
- the `Invalidate` method is called
- it is a global (zrepl process-wide) cache
fixes#316
package trace:
- introduce the concept of tasks and spans, tracked as linked list within ctx
- see package-level docs for an overview of the concepts
- **main feature 1**: unique stack of task and span IDs
- makes it easy to follow a series of log entries in concurrent code
- **main feature 2**: ability to produce a chrome://tracing-compatible trace file
- either via an env variable or a `zrepl pprof` subcommand
- this is not a CPU profile, we already have go pprof for that
- but it is very useful to visually inspect where the
replication / snapshotter / pruner spends its time
( fixes#307 )
usage in package daemon/logging:
- goal: every log entry should have a trace field with the ID stack from package trace
- make `logging.GetLogger(ctx, Subsys)` the authoritative `logger.Logger` factory function
- the context carries a linked list of injected fields which
`logging.GetLogger` adds to the logger it returns
- `logging.GetLogger` also uses package `trace` to get the
task-and-span-stack and injects it into the returned logger's fields
- **Resumable Send & Recv Support**
No knobs required, automatically used where supported.
- **Hold-Protected Send & Recv**
Automatic ZFS holds to ensure that we can always resume a replication step.
- **Encrypted Send & Recv Support** for OpenZFS native encryption.
Configurable at the job level, i.e., for all filesystems a job is responsible for.
- **Receive-side hold on last received dataset**
The counterpart to the replication cursor bookmark on the send-side.
Ensures that incremental replication will always be possible between a sender and receiver.
Design Doc
----------
`replication/design.md` doc describes how we use ZFS holds and bookmarks to ensure that a single replication step is always resumable.
The replication algorithm described in the design doc introduces the notion of job IDs (please read the details on this design doc).
We reuse the job names for job IDs and use `JobID` type to ensure that a job name can be embedded into hold tags, bookmark names, etc.
This might BREAK CONFIG on upgrade.
Protocol Version Bump
---------------------
This commit makes backwards-incompatible changes to the replication/pdu protobufs.
Thus, bump the version number used in the protocol handshake.
Replication Cursor Format Change
--------------------------------
The new replication cursor bookmark format is: `#zrepl_CURSOR_G_${this.GUID}_J_${jobid}`
Including the GUID enables transaction-safe moving-forward of the cursor.
Including the job id enables that multiple sending jobs can send the same filesystem without interfering.
The `zrepl migrate replication-cursor:v1-v2` subcommand can be used to safely destroy old-format cursors once zrepl has created new-format cursors.
Changes in This Commit
----------------------
- package zfs
- infrastructure for holds
- infrastructure for resume token decoding
- implement a variant of OpenZFS's `entity_namecheck` and use it for validation in new code
- ZFSSendArgs to specify a ZFS send operation
- validation code protects against malicious resume tokens by checking that the token encodes the same send parameters that the send-side would use if no resume token were available (i.e. same filesystem, `fromguid`, `toguid`)
- RecvOptions support for `recv -s` flag
- convert a bunch of ZFS operations to be idempotent
- achieved through more differentiated error message scraping / additional pre-/post-checks
- package replication/pdu
- add field for encryption to send request messages
- add fields for resume handling to send & recv request messages
- receive requests now contain `FilesystemVersion To` in addition to the filesystem into which the stream should be `recv`d into
- can use `zfs recv $root_fs/$client_id/path/to/dataset@${To.Name}`, which enables additional validation after recv (i.e. whether `To.Guid` matched what we received in the stream)
- used to set `last-received-hold`
- package replication/logic
- introduce `PlannerPolicy` struct, currently only used to configure whether encrypted sends should be requested from the sender
- integrate encryption and resume token support into `Step` struct
- package endpoint
- move the concepts that endpoint builds on top of ZFS to a single file `endpoint/endpoint_zfs.go`
- step-holds + step-bookmarks
- last-received-hold
- new replication cursor + old replication cursor compat code
- adjust `endpoint/endpoint.go` handlers for
- encryption
- resumability
- new replication cursor
- last-received-hold
- client subcommand `zrepl holds list`: list all holds and hold-like bookmarks that zrepl thinks belong to it
- client subcommand `zrepl migrate replication-cursor:v1-v2`
ActiveSide.do() can only run sequentially, i.e. we cannot run
replication and pruning in parallel. Why?
* go-streamrpc only allows one active request at a time
(this is bad design and should be fixed at some point)
* replication and pruning are implemented independently, but work on the
same resources (snapshots)
A: pruning might destroy a snapshot that is planned to be replicated
B: replication might replicate snapshots that should be pruned
We do not have any resource management / locking for A and B, but we
have a use case where users don't want their machine fill up with
snapshots if replication does not work.
That means we _have_ to run the pruners.
A further complication is that we cannot just cancel the replication
context after a timeout and move on to the pruner: it could be initial
replication and we don't know how long it will take.
(And we don't have resumable send & recv yet).
With the previous commits, we can implement the watchdog using context
cancellation.
Note that the 'MadeProgress()' calls can only be placed right before
non-error state transition. Otherwise, we could end up in a live-lock.
We have the problem that there are legitimate use cases where a user
does not want their machine to fill up with snapshots, even if it means
unreplicated must be destroyed. This can be expressed by *not*
configuring the keep rule `not_replicated` for the snapshot-creating
side. This commit only addresses push mode because we don't support
pruning in the source job. We adivse users in the docs to use push mode
if they have above use case, so this is fine - at least for 0.1.
Ideally, the replication.Replication would communicate to the pruner
which snapshots are currently part of the replication plan, and then
we'd need some conflict resolution to determine whether it's more
important to destroy the snapshots or to replicate them (destroy should
win?).
However, we don't have the infrastructure for this yet (we could parse
the replication report, but that's just ugly). And we want to get 0.1
out, so showtime for a dirty hack:
We start replication, and ideally, replication and pruning is done
before new snapshot have been taken. If so: great. However, what happens
if snapshots have been taken and we are not done with replication and /
or pruning?
* If replicatoin is making progress according to its state, let it run.
This covers the *important* situation of initial replication, where
replication may easily take longer than a single snapshotting interval.
* If replication is in an error state, cancel it through context
cancellation.
* As with the pruner below, the main problem here is that
status output will only contain "context cancelled" after the
cancellation, instead of showing the reason why it was cancelled.
Not nice, but oh well, the logs provide enough detail for this
niche situation...
* If we are past replication, we're still pruning
* Leave the local (send-side) pruning alone.
Again, we only implement this hack for push, so we know sender is
local, and it will only fail hard, not retry.
* If the remote (receiver-side) pruner is in an error state, cancel it
through context cancellation.
* Otherwise, let it run.
Note that every time we "let it run", we tolerate a temporary excess of
snapshots, but given sufficiently aggressive timeouts and the assumption
that the snapshot interval is much greater than the timeouts, this is
not a significant problem in practice.
If concurrent updates strictly modify *different* members of the tasks
struct, the copying + lock-drop still constitutes a race condition:
The last updater always wins and sets tasks to its copy + changes.
This eliminates the other updater's changes.
