## Problem
When Availabilities are created, the amount of bytes in the Availability are reserved in the repo, so those bytes on disk cannot be written to otherwise. When a request for storage is received by a node, if a previously created Availability is matched, an attempt will be made to fill a slot in the request (more accurately, the request's slots are added to the SlotQueue, and eventually those slots will be processed). During download, bytes that were reserved for the Availability were released (as they were written to disk). To prevent more bytes from being released than were reserved in the Availability, the Availability was marked as used during the download, so that no other requests would match the Availability, and therefore no new downloads (and byte releases) would begin. The unfortunate downside to this, is that the number of Availabilities a node has determines the download concurrency capacity. If, for example, a node creates a single Availability that covers all available disk space the operator is willing to use, that single Availability would mean that only one download could occur at a time, meaning the node could potentially miss out on storage opportunities.
## Solution
To alleviate the concurrency issue, each time a slot is processed, a Reservation is created, which takes size (aka reserved bytes) away from the Availability and stores them in the Reservation object. This can be done as many times as needed as long as there are enough bytes remaining in the Availability. Therefore, concurrent downloads are no longer limited by the number of Availabilities. Instead, they would more likely be limited to the SlotQueue's `maxWorkers`.
From a database design perspective, an Availability has zero or more Reservations.
Reservations are persisted in the RepoStore's metadata, along with Availabilities. The metadata store key path for Reservations is ` meta / sales / reservations / <availabilityId> / <reservationId>`, while Availabilities are stored one level up, eg `meta / sales / reservations / <availabilityId> `, allowing all Reservations for an Availability to be queried (this is not currently needed, but may be useful when work to restore Availability size is implemented, more on this later).
### Lifecycle
When a reservation is created, its size is deducted from the Availability, and when a reservation is deleted, any remaining size (bytes not written to disk) is returned to the Availability. If the request finishes, is cancelled (expired), or an error occurs, the Reservation is deleted (and any undownloaded bytes returned to the Availability). In addition, when the Sales module starts, any Reservations that are not actively being used in a filled slot, are deleted.
Having a Reservation persisted until after a storage request is completed, will allow for the originally set Availability size to be reclaimed once a request contract has been completed. This is a feature that is yet to be implemented, however the work in this PR is a step in the direction towards enabling this.
### Unknowns
Reservation size is determined by the `StorageAsk.slotSize`. If during download, more bytes than `slotSize` are attempted to be downloaded than this, then the Reservation update will fail, and the state machine will move to a `SaleErrored` state, deleting the Reservation. This will likely prevent the slot from being filled.
### Notes
Based on #514
## Slot queue
Adds a slot queue, as per the [slot queue design](https://github.com/codex-storage/codex-research/blob/master/design/sales.md#slot-queue).
Any time storage is requested, all slots from that request are immediately added to the queue. Finished, Canclled, Failed requests remove all slots with that request id from the queue. SlotFreed events add a new slot to the queue and SlotFilled events remove the slot from the queue. This allows popping of a slot each time one is processed, making things much simpler.
When an entire request of slots is added to the queue, the slot indices are shuffled randomly to hopefully prevent nodes that pick up the same storage requested event from clashing on the first processed slot index. This allowed removal of assigning a random slot index in the SalePreparing state and it also ensured that all SalesAgents will have a slot index assigned to them at the start thus the removal of the optional slotIndex.
Remove slotId from SlotFreed event as it was not being used. RequestId and slotIndex were added to the SlotFreed event earlier and those are now being used
The slot queue invariant that prioritises queue items added to the queue relies on a scoring mechanism to sort them based on the [sort order in the design document](https://github.com/codex-storage/codex-research/blob/master/design/sales.md#sort-order).
When a storage request is handled by the sales module, a slot index was randomly assigned and then the slot was filled. Now, a random slot index is only assigned when adding an entire request to the slot queue. Additionally, the slot is checked that its state is `SlotState.Free` before continuing with the download process.
SlotQueue should always ensure the underlying AsyncHeapQueue has one less than the maximum items, ensuring the SlotQueue can always have space to add an additional item regardless if it’s full or not.
Constructing `SlotQueue.workers` in `SlotQueue.new` calls `newAsyncQueue` which causes side effects, so the construction call had to be moved to `SlotQueue.start`.
Prevent loading request from contract (network request) if there is an existing item in queue for that request.
Check availability before adding request to queue.
Add ability to query market contract for past events. When new availabilities are added, the `onReservationAdded` callback is triggered in which past `StorageRequested` events are queried, and those slots are added to the queue (filtered by availability on `push` and filtered by state in `SalePreparing`).
#### Request Workers
Limit the concurrent requests being processed in the queue by using a limited pool of workers (default = 3). Workers are in a data structure of type `AsyncQueue[SlotQueueWorker]`. This allows us to await a `popFirst` for available workers inside of the main SlotQueue event loop
Add an `onCleanUp` that stops the agents and removes them from the sales module agent list. `onCleanUp` is called from sales end states (eg ignored, cancelled, finished, failed, errored).
Add a `doneProcessing` future to `SlotQueueWorker` to be completed in the `OnProcessSlot` callback. Each `doneProcessing` future created is cancelled and awaited in `SlotQueue.stop` (thanks to `TrackableFuturees`), which forced `stop` to become async.
- Cancel dispatched workers and the `onProcessSlot` callbacks, prevents zombie callbacks
#### Add TrackableFutures
Allow tracking of futures in a module so they can be cancelled at a later time. Useful for asyncSpawned futures, but works for any future.
### Sales module
The sales module needed to subscribe to request events to ensure that the request queue was managed correctly on each event. In the process of doing this, the sales agents were updated to avoid subscribing to events in each agent, and instead dispatch received events from the sales module to all created sales agents. This would prevent memory leaks on having too many eventemitters subscribed to.
- prevent removal of agents from sales module while stopping, otherwise the agents seq len is modified while iterating
An additional sales agent state was added, `SalePreparing`, that handles all state machine setup, such as retrieving the request and subscribing to events that were previously in the `SaleDownloading` state.
Once agents have parked in an end state (eg ignored, cancelled, finished, failed, errored), they were not getting cleaned up and the sales module was keeping a handle on their reference. An `onCleanUp` callback was created to be called after the state machine enters an end state, which could prevent a memory leak if the number of requests coming in is high.
Move the SalesAgent callback raises pragmas from the Sales module to the proc definition in SalesAgent. This avoids having to catch `Exception`.
- remove unneeded error handling as pragmas were moved
Move sales.subscriptions from an object containing named subscriptions to a `seq[Subscription]` directly on the sales object.
Sales tests: shut down repo after sales stop, to fix SIGABRT in CI
### Add async Promise API
- modelled after JavaScript Promise API
- alternative to `asyncSpawn` that allows handling of async calls in a synchronous context (including access to the synchronous closure) with less additional procs to be declared
- Write less code, catch errors that would otherwise defect in asyncspawn, and execute a callback after completion
- Add cancellation callbacks to utils/then, ensuring cancellations are handled properly
## Dependencies
- bump codex-contracts-eth to support slot queue (https://github.com/codex-storage/codex-contracts-eth/pull/61)
- bump nim-ethers to 0.5.0
- Bump nim-json-rpc submodule to 0bf2bcb
---------
Co-authored-by: Jaremy Creechley <creechley@gmail.com>
* [marketplace] reservations module
- add de/serialization for Availability
- add markUsed/markUnused in persisted availability
- add query for unused
- add reserve/release
- reservation module tests
- split ContractInteractions into client contracts and host contracts
- remove reservations start/stop as the repo start/stop is being managed by the node
- remove dedicated reservations metadata store and use the metadata store from the repo instead
- Split ContractInteractions into:
- ClientInteractions (with purchasing)
- HostInteractions (with sales and proving)
- compilation fix for nim 1.2
[repostore] fix started flag, add tests
[marketplace] persist slot index
For loading the sales state from chain, the slot index was not previously persisted in the contract. Will retrieve the slot index from the contract when the sales state is loaded.
* Revert repostore changes
In favour of separate PR https://github.com/status-im/nim-codex/pull/374.
* remove warnings
* clean up
* tests: stop repostore during teardown
* change constructor type identifier
Change Contracts constructor to accept Contracts type instead of ContractInteractions.
* change constructor return type to Result instead of Option
* fix and split interactions tests
* clean up, fix tests
* find availability by slot id
* remove duplication in host/client interactions
* add test for finding availability by slotId
* log instead of raiseAssert when failed to mark availability as unused
* move to SaleErrored state instead of raiseAssert
* remove unneeded reverse
It appears that order is not preserved in the repostore, so reversing does not have the intended effect here.
* update open api spec for potential rest endpoint errors
* move functions about available bytes to repostore
* WIP: reserve and release availabilities as needed
WIP: not tested yet
Availabilities are marked as used when matched (just before downloading starts) so that future matching logic does not match an availability currently in use.
As the download progresses, batches of blocks are written to disk, and the equivalent bytes are released from the reservation module. The size of the availability is reduced as well.
During a reserve or release operation, availability updates occur after the repo is manipulated. If the availability update operation fails, the reserve or release is rolled back to maintain correct accounting of bytes.
Finally, once download completes, or if an error occurs, the availability is marked as unused so future matching can occur.
* delete availability when all bytes released
* fix tests + cleanup
* remove availability from SalesContext callbacks
Availability is no longer used past the SaleDownloading state in the state machine. Cleanup of Availability (marking unused) is handled directly in the SaleDownloading state, and no longer in SaleErrored or SaleFinished. Likewise, Availabilities shouldn’t need to be handled on node restart.
Additionally, Availability was being passed in SalesContext callbacks, and now that Availability is only used temporarily in the SaleDownloading state, Availability is contextually irrelevant to the callbacks, except in OnStore possibly, though it was not being consumed.
* test clean up
* - remove availability from callbacks and constructors from previous commit that needed to be removed (oopsie)
- fix integration test that checks availabilities
- there was a bug fixed that crashed the node due to a missing `return success` in onStore
- the test was fixed by ensuring that availabilities are remaining on the node, and the size has been reduced
- change Availability back to non-ref object and constructor back to init
- add trace logging of all state transitions in state machine
- add generally useful trace logging
* fixes after rebase
1. Fix onProve callbacks
2. Use Slot type instead of tuple for retrieving active slot.
3. Bump codex-contracts-eth that exposes getActivceSlot call.
* swap contracts branch to not support slot collateral
Slot collateral changes in the contracts require further changes in the client code, so we’ll skip those changes for now and add in a separate commit.
* modify Interactions and Deployment constructors
- `HostInteractions` and `ClientInteractions` constructors were simplified to take a contract address and no overloads
- `Interactions` prepared simplified so there are no overloads
- `Deployment` constructor updated so that it takes an optional string parameter, instead `Option[string]`
* Move `batchProc` declaration
`batchProc` needs to be consumed by both `node` and `salescontext`, and they can’t reference each other as it creates a circular dependency.
* [reservations] rename `available` to `hasAvailable`
* [reservations] default error message to inner error msg
* add SaleIngored state
When a storage request is handled but the request does match availabilities, the sales agent machine is sent to the SaleIgnored state. In addition, the agent is constructed in a way that if the request is ignored, the sales agent is removed from the list of active agents being tracked in the sales module.
* [marketplace] get active slots from chain
# Conflicts:
# codex/contracts/market.nim
* [marketplace] make on chain event callbacks async
# Conflicts:
# tests/codex/helpers/mockmarket.nim
* [marketplace] make availability optional for node restart
# Conflicts:
# tests/codex/testsales.nim
* [marketplace] add async state machine
Allows for `enterAsync` to be cancelled.
* [marketplace] move sale process to async state machine
* [marketplace] sales state machine tests
* bump dagger-contracts
* [marketplace] fix ci issue with chronicles output
* PR comments
- add slotIndex to `SalesAgent` constructor
- remove `SalesAgent.init`
- rename `SalesAgent.init` to `start` and `SalesAgent.deinit` to `stop`.
- rename `SalesAgent. populateRequest` to `SalesAgent.retreiveRequest`.
- move availability removal to the downloading state. once availability is persisted to disk, it should survive node restarts.
-
* [marketplace] handle slot filled by other host
Handle the case when in the downloading, proving, or filling states, that another host fills the slot.
* [marketplace] use requestId for mySlots
* [marketplace] infer slot index from slotid
prevents reassigning a random slot index when restoring state from chain
* [marketplace] update to work with latest contracts
* [marketplace] clean up
* [marketplace] align with contract changes
- getState / state > requestState
- getSlot > getRequestFromSlotId
- support MarketplaceConfig
- support slotState, remove unneeded Slot type
- collateral > config.collateral.initialAmount
- remove proofPeriod contract call
- Revert reason “Slot empty” > “Slot is free”
- getProofEnd > read SlotState
Tests for changes
* [marketplace] add missing file
* [marketplace] bump codex-contracts-eth
* [config] remove unused imports
* [sales] cleanup
* [sales] fix: do not crash when fetching state fails
* [sales] make slotIndex non-optional
* Rebase and update NBS commit
Rebase on top of main and update NBS commit to the CI fix.
* [marketplace] use async subscription event handlers
* [marketplace] support slotIndex no longer optional
Previously, SalesAgent.slotIndex had been moved to not optional. However, there were still many places where optionality was assumed. This commit removes those assumuptions.
* [marketplace] sales state machine: use slotState
Use `slotState` instead of `requestState` for sales state machine.
* [marketplace] clean up
* [statemachine] adds a statemachine for async workflows
Allows events to be scheduled synchronously.
See https://github.com/status-im/nim-codex/pull/344
Co-Authored-By: Ben Bierens <thatbenbierens@gmail.com>
Co-Authored-By: Eric Mastro <eric.mastro@gmail.com>
* [market] make market callbacks synchronous
* [statemachine] export Event
* [statemachine] ensure that no errors are raised
* [statemachine] add machine parameter to run method
* [statemachine] initialize queue on start
* [statemachine] check futures before cancelling them
* [sales] use new async state machine
- states use new run() method and event mechanism
- StartState starts subscriptions and loads request
* [statemachine] fix unsusbscribe before subscribe
* [sales] replace old state transition tests
* [sales] separate state machine from sales data
* [sales] remove reference from SalesData to Sales
* [sales] separate sales context from sales
* [sales] move decoupled types into their own modules
* [sales] move retrieveRequest to SalesData
* [sales] move subscription logic into SalesAgent
* [sales] unsubscribe when finished or errored
* [build] revert back to released version of nim-ethers
* [sales] remove SaleStart state
* [sales] add missing base method
* [sales] move asyncSpawn helper to utils
* [sales] fix imports
* [sales] remove unused variables
* [sales statemachine] add async state machine error handling (#349)
* [statemachine] add error handling to asyncstatemachine
- add error handling to catch errors during state.run
- Sales: add ErrorState to identify which state to transition to during an error. This had to be added to SalesAgent constructor due to circular dependency issues, otherwise it would have been added directly to SalesAgent.
- Sales: when an error during run is encountered, the SaleErrorState is constructed with the error, and by default (base impl) will return the error state, so the machine can transition to it. This can be overridden by individual states if needed.
* [sales] rename onSaleFailed to onSaleErrored
Because there is already a state named SaleFailed which is meant to react to an onchain RequestFailed event and also because this callback is called from SaleErrored, renaming to onSaleErrored prevents ambiguity and confusion as to what has happened at the callback callsite.
* [statemachine] forward error to state directly
without going through a machine method first
* [statemachine] remove unnecessary error handling
AsyncQueueFullError is already handled in schedule()
* [statemachine] test that cancellation ignores onError
* [sales] simplify error handling in states
Rely on the state machine error handling instead
of catching errors in the state run method
---------
Co-authored-by: Mark Spanbroek <mark@spanbroek.net>
* [statemachine] prevent memory leaks
prevent memory leaks and nil access defects by:
- allowing multiple subscribe/unsubscribes of salesagent
- disallowing individual salesagent subscription calls to be made externally (requires the .subscribed check)
- allowing mutiple start/stops of asyncstatemachine
- disregard asyncstatemachine schedules if machine not yet started
* [salesagent] add salesagent-specific tests
1. test multiple subscribe/unsubscribes
2. test scheduling machine without being started
3. test subscriptions are working correctly with external events
4. test errors can be overridden at the state level for ErrorHandlingStates.
---------
Co-authored-by: Eric Mastro <eric.mastro@gmail.com>
Co-authored-by: Mark Spanbroek <mark@spanbroek.net>
Co-authored-by: Ben Bierens <thatbenbierens@gmail.com>
Ben Bierens
Adam Uhlíř
Eric
Eric Mastro