- Promote a data-structure-store design paradigm: build your applications out of known data types so that the system can reason about commutativity, replication, etc, and you can use strict consistency
- Use commutativity to avoid falling over from contention on bursty, skewed, update-heavy workloads
- i.e. Ellen retweet, Truffle auction, Reddit
- Apply abstract locks (from database and multi-core STM communities) to distributed ad-hoc applications like those that use Redis
- Faster, better, linearizable, transactional Redis.
- to apply ideas from combining and abstract locks to distributed transactions
- avoid hotspots automatically by splitting/replicating on the fly
- apply a more rigorous notion of serializability and consistency on data structures
- more generalized notion of commutative ops make transactions more efficient (allow them to commit even while operating on replicas, more concurrency)
- better programming model that allows complex operations to be efficiently/scalably executed
- other systems require inventing new operators with different semantics, or splitting out into multiple records/keys in database
- e.g. keep track of max bid, separate from who did the bid, from the total number of bids. if it was just a set, it would be easier to think about
- asynchronous phasing/reconciliation will perform better than global bulk-synchronous approaches
- exposing more mergeable operations will allow more concurrency and therefore greater performance
- combining will reduce the overhead (processing time & data movement) of staging and reordering transactions
- data structures = better programming model
- combining operations in transactions
- generalization of phases: not just commutative and non-commutative, but which operations commute with other ops (also generalizes "commutative-requests-first" transaction reordering)
- generalizing "key-splitting" (proxy/replicas) for any supported data structure
- How to do splitting?
- Complete replicas: Keep track of operations, combine them into more efficient update, and send to all replicas
- don't have to collapse replicas just to do ops that need global state
- after sharing, everyone has the complete "log", so we can service all requests locally
- is this too much more communication? seems like we can expect that hot records will continue to be hot, so we should keep them split...
- all replicas have to know about all other replicas... (maybe we can do as a tree?)
- Splits: Essentially "empty" versions of original data structure, with defined semantics for merging into the original
- could be done with a Monoid-like interface
- limited operations, anything requiring concrete state must wait
- Proxies: Custom logic for how to take in operations, combine them, and spit out an answer.
- allows for annihilation
- saves from sending state of large object at splitting time
- data structure maintainer must implement by hand?
- Asynchronous record/object reconciliations or global phases?
- Seems like most of the current work I've seen relies on global phases (both OSDI'14 papers at least). My intuition is that any time you implement a global synchronization, you're probably leaving performance on the floor. Are they doing this because it's an easier implementation? Maybe in the presence of coordinating multi-object transactions, it's too hard to do asynchronously and you just default to global phases?
- Combining with transactions
- combined ops (patches) include a set of transactions, but 1 transaction may be split among multiple patches
- is there any chance of having unserializable sets of combined ops? (for example, can I end up with some operations in one patch that rely on a particular order, which is different than the order observed for the second patch?)
- I'm thinking that this shouldn't be a problem since we're only combining ops that commute w.r.t the current state/mode of the object, but because we're splitting transactions and recombining them per-object this is worth considering carefully.
- how do we efficiently keep track of which transactions are in a given patch? how do we know when it's been completed?
- using the notion of futures (returned by atomic ops) may make this easier to think about.
- can allow futures to be chained into multi-hop transactions
- individual futures can be combined locally into more compact multi-futures
- Would the resulting protocol be too complex?
- Is it weird to work on it while in submission?
- Seems like there is some synergy in that there are already multiple replicas in memory to work on.
- Reads typically just need to hit one replica; maybe commutative/combining ops can do the same?
- Do combining at replicas; when sharing results with others, apply them in a deterministic order (or if all commute, then whatever order you want)
First pass: add Redis data structure ops to increase concurrency / reduce aborts
- local backing store for each shard/replica
- currently a KV store (put/get)
- todo: extend/replace with Redis (or something like Redis)
- OCC layer
- before things get put in the backing store, you stage them through the OCC layer, which tracks read/write sets of all in-flight txs
prepare, you add to the read/write sets
commit, put them into the backing store and remove them from the read/write sets
- use timestamps to determine if something conflicts.
- can do some reordering because you can sometimes choose the timestamp so that it works
- extend this to allow commutative ops to not abort; need something more than just read/write sets.
- could also try doing it with locks...
- use timestamps to determine commutative phases
- zeroth pass: disable txn/serializability by just returning
true all the time (then you basically get a form of distributed Redis)
- run under Tapir or just traditional Paxos (or both)
- try to stick all the higher-level operations into
put ops as far as the upper layers (Tapir/Paxos) are concerned. the OCC layer can sort out what the operations actually are (given the type of the object) and determine if they conflict or not
OCC or Locking?
- use OCC when the vast majority of cases will succeed. this is the case even for contentious workloads if all the contending transactions commute (so there must not be other transactions that read and modify the same record)
- it seemed like there might be a way to cheat a bit by allowing transactions that only read the contended record and do other independent things (e.g.
post), because we could just have them read the current un-committed value & timestamp
- this seems fine unless the order of two of these txns can be observed. the case we had was two posts, where P1 saw a later committed version of the follower list than P2, but ended up being ordered before P2 in the timeline because we claimed that the timeline is a set ordered by post timestamp so insertions commute. but it seems like they don't commute because of the link to actual real time
- locks make it easier to support mixed workloads
- we can more easily implement phases where non-commuting ops block while ops that commute with the current lock state can proceed immediately
- pay the cost of checking every time (though OCC also has to do checks...)
- need a sole coordinator for the lock (or maybe not if my consistent split replicas work)
- can't concurrently read from the previous snapshot (is this true? if we haven't committed them yet, reads could still be done)
- still can't perform side effects until we've acquired all the locks -- so things will still need to be split-phase
- is starvation more of a problem in one or the other?
My OCC Protocol
- Execute transaction: do pure 'stage' step for each op; each 'stage':
- may trivially return (e.g. for simple 'init' or 'set' ops)
- or optimistically continue, even though a conflict is possible (e.g. 'read' op may be able to be put before commutative ops)
- or immediately abort txn (if it's clear it will never work out)
- Prepare transaction
- check that all ops in txn are allowable (e.g. all commute, or if doing 'read', must be completely independent)
- Commit transaction
- execute 'apply' action for each op in txn (performs some kind of mutation)
Batching & combining
- add a level of batching to the system that allows you to amortize the checks over commutative phases
- expose the fact that you're combining multiple txns into a single message to the upper layers
- try to come up with a protocol for combining over replicas
Contention avoidance by splitting
- split hot keys across multiple shards
- in tapir: consensus then has to happen first across the replicas of each split, then over the combined result
- in paxos: the leaders of each shard's replica groups have to elect a leader, who then combines all of their results
- this adds many extra round trips, probably only will pay off if you have highly contentious keys so you can batch many transactions and do the extra layer of coordination infrequently
- first evaluate this on Grappa or a non-fault-tolerant (or non-replicated) platform and see what the performance tradeoffs are for splitting, given particular workloads
- then try the fault-tolerant version with higher cost for synchronizing
- dynamic splitting
- use consistent hashing, plus a replica number (
k => hash(k)+hash(replica_idx) % nnode)
- everyone just needs to know how many replicas a given key has (or a lower bound on it) so they can choose from the set of hashed locations
Example: Counter (not to be confused with counterexample)
- "Likes" on social network pages
read: => T
write: => void
incr: => void
fetch_add: => T (??)
incr >> read (or