ChallengeEfficient snapshotable state
At the heart of RavenDB, there is a data structure that we call the Page Translation Table. It is one of the most important pieces inside RavenDB.
The page translation table is basically a Dictionary<long, Page>, mapping between a page number and the actual page. The critical aspect of this data structure is that it is both concurrent and multi-version. That is, at a single point, there may be multiple versions of the table, representing different versions of the table at given points in time.
The way it works, a transaction in RavenDB generates a page translation table as part of its execution and publishes the table on commit. However, each subsequent table builds upon the previous one, so things become more complex. Here is a usage example (in Python pseudo-code):
table = {}
with wtx1 = write_tx(table):
wtx1.put(2, 'v1')
wtx1.put(3, 'v1')
wtx1.publish(table)
# table has (2 => v1, 3 => v1)
with wtx2 = write_tx(table):
wtx2.put(2, 'v2')
wtx2.put(4, 'v2')
wtx2.publish(table)
# table has (2 => v2, 3 => v1, 4 => v2)
This is pretty easy to follow, I think. The table is a simple hash table at this point in time.
The catch is when we mix read transactions as well, like so:
# table has (2 => v2, 3 => v1, 4 => v2)
with rtx1 = read_tx(table):
with wtx3 = write_tx(table):
wtx3.put(2, 'v3')
wtx3.put(3, 'v3')
wtx3.put(5, 'v3')
with rtx2 = read_tx(table):
rtx2.read(2) # => gives, v2
rtx2.read(3) # => gives, v1
rtx2.read(5) # => gives, None
wtx3.publish(table)
# table has (2 => v3, 3 => v3, 4 => v2, 5 => v3)
# but rtx2 still observe the value as they were when
# rtx2 was created
rtx2.read(2) # => gives, v2
rtx2.read(3) # => gives, v1
rtx2.read(5) # => gives, None
In other words, until we publish a transaction, its changes don’t take effect. And any read translation that was already started isn’t impacted. We also need this to be concurrent, so we can use the table in multiple threads (a single write transaction at a time, but potentially many read transactions). Each transaction may modify hundreds or thousands of pages, and we’ll only clear the table of old values once in a while (so it isn’t infinite growth, but may certainly reach respectable numbers of items).
The implementation we have inside of RavenDB for this is complex! I tried drawing that on the whiteboard to explain what was going on, and I needed both the third and fourth dimensions to illustrate the concept.
Given these requirements, how would you implement this sort of data structure?
More posts in "Challenge" series:
- (01 Jul 2024) Efficient snapshotable state
- (13 Oct 2023) Fastest node selection metastable error state–answer
- (12 Oct 2023) Fastest node selection metastable error state
- (19 Sep 2023) Spot the bug
- (04 Jan 2023) what does this code print?
- (14 Dec 2022) What does this code print?
- (01 Jul 2022) Find the stack smash bug… – answer
- (30 Jun 2022) Find the stack smash bug…
- (03 Jun 2022) Spot the data corruption
- (06 May 2022) Spot the optimization–solution
- (05 May 2022) Spot the optimization
- (06 Apr 2022) Why is this code broken?
- (16 Dec 2021) Find the slow down–answer
- (15 Dec 2021) Find the slow down
- (03 Nov 2021) The code review bug that gives me nightmares–The fix
- (02 Nov 2021) The code review bug that gives me nightmares–the issue
- (01 Nov 2021) The code review bug that gives me nightmares
- (16 Jun 2021) Detecting livelihood in a distributed cluster
- (21 Apr 2020) Generate matching shard id–answer
- (20 Apr 2020) Generate matching shard id
- (02 Jan 2020) Spot the bug in the stream
- (28 Sep 2018) The loop that leaks–Answer
- (27 Sep 2018) The loop that leaks
- (03 Apr 2018) The invisible concurrency bug–Answer
- (02 Apr 2018) The invisible concurrency bug
- (31 Jan 2018) Find the bug in the fix–answer
- (30 Jan 2018) Find the bug in the fix
- (19 Jan 2017) What does this code do?
- (26 Jul 2016) The race condition in the TCP stack, answer
- (25 Jul 2016) The race condition in the TCP stack
- (28 Apr 2015) What is the meaning of this change?
- (26 Sep 2013) Spot the bug
- (27 May 2013) The problem of locking down tasks…
- (17 Oct 2011) Minimum number of round trips
- (23 Aug 2011) Recent Comments with Future Posts
- (02 Aug 2011) Modifying execution approaches
- (29 Apr 2011) Stop the leaks
- (23 Dec 2010) This code should never hit production
- (17 Dec 2010) Your own ThreadLocal
- (03 Dec 2010) Querying relative information with RavenDB
- (29 Jun 2010) Find the bug
- (23 Jun 2010) Dynamically dynamic
- (28 Apr 2010) What killed the application?
- (19 Mar 2010) What does this code do?
- (04 Mar 2010) Robust enumeration over external code
- (16 Feb 2010) Premature optimization, and all of that…
- (12 Feb 2010) Efficient querying
- (10 Feb 2010) Find the resource leak
- (21 Oct 2009) Can you spot the bug?
- (18 Oct 2009) Why is this wrong?
- (17 Oct 2009) Write the check in comment
- (15 Sep 2009) NH Prof Exporting Reports
- (02 Sep 2009) The lazy loaded inheritance many to one association OR/M conundrum
- (01 Sep 2009) Why isn’t select broken?
- (06 Aug 2009) Find the bug fixes
- (26 May 2009) Find the bug
- (14 May 2009) multi threaded test failure
- (11 May 2009) The regex that doesn’t match
- (24 Mar 2009) probability based selection
- (13 Mar 2009) C# Rewriting
- (18 Feb 2009) write a self extracting program
- (04 Sep 2008) Don't stop with the first DSL abstraction
- (02 Aug 2008) What is the problem?
- (28 Jul 2008) What does this code do?
- (26 Jul 2008) Find the bug fix
- (05 Jul 2008) Find the deadlock
- (03 Jul 2008) Find the bug
- (02 Jul 2008) What is wrong with this code
- (05 Jun 2008) why did the tests fail?
- (27 May 2008) Striving for better syntax
- (13 Apr 2008) calling generics without the generic type
- (12 Apr 2008) The directory tree
- (24 Mar 2008) Find the version
- (21 Jan 2008) Strongly typing weakly typed code
- (28 Jun 2007) Windsor Null Object Dependency Facility
Comments
The runtime's
Dictionary<TKey, TValue>
is implemented using a hash lookup for the head of a linked list, and then iterating through the linked list to find an exact key match.This design is perfect for extending to a single-writer multiple reader scenario. Just add a created/deleted version number to each entry in the linked list and the rest remains the same. When taking a snapshot, re-use the same heads / entries - then to lookup an element, only consider elements created before the current version and deleted after.
To add an element when the dictionary is full, then do clean-up and permanently remove deleted items, reset version number, etc since the new heads/entries are not yet used.
To add an element to existing head/entries arrays, first add the entry with the new version number and then update the heads to point to this new entry. Existing snapshots will either skip it due to a higher version number, or skip it because the head hasn't yet been updated. Either way the same result.
I've put this in a GitHub gist.
Benchmarks show performance comparable to
Dictionary<TKey, TValue>
for construction and access although a bit slower due of the various micro-optimizations in the built-in implementation.Please feel free to take the code and adapt it for your needs commercial or otherwise. In fact I'm thinking to package it in a small utility library if there's value to it.
Comment preview