Primary caching 8: implement latest-at data-time cache entry deduplication #4712
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
teh-cmc
added
🔍 re_query
teh-cmc
changed the title
This was referenced Jan 5, 2024
teh-cmc
force-pushed
the
cmc/primcache_7_data_time
branch
from
bcc7ab9 to
035a111
Compare
teh-cmc
force-pushed
the
cmc/primcache_8_dedupe
branch
from
4d4418c to
31e55d2
Compare
teh-cmc
commented
Jan 8, 2024
teh-cmc
force-pushed
the
cmc/primcache_8_dedupe
branch
from
e080a99 to
1517ec4
Compare
teh-cmc
force-pushed
the
cmc/primcache_7_data_time
branch
from
64f9cf4 to
88989df
Compare
teh-cmc
force-pushed
the
cmc/primcache_8_dedupe
branch
from
1517ec4 to
317c0ec
Compare
4 tasks
teh-cmc
added a commit
that referenced
this pull request
Jan 10, 2024
This implements the most barebone latest-at caching support. The goal is merely to introduce all the machinery and boilerplate required to get the primary cache running, actual caching features will be implemented on top of this foundation in follow up PRs. The [existing benchmark suite](https://github.com/rerun-io/rerun/blob/790f391/crates/re_query/benches/query_benchmark.rs) has been ported as-is to the cached APIs (5950X, Arch): ``` group primcache_3_vanilla primcache_3_cached ----- ------------------- ------------------ arrow_batch_points2/insert 1.02 1015.0±11.07µs 939.6 MElem/sec 1.00 1000.0±7.37µs 953.7 MElem/sec arrow_batch_points2/query 2.90 3.4±0.02µs 276.5 MElem/sec 1.00 1190.5±41.55ns 801.0 MElem/sec arrow_batch_strings2/insert 1.00 1045.7±7.85µs 912.0 MElem/sec 1.00 1042.1±14.01µs 915.1 MElem/sec arrow_batch_strings2/query 1.91 21.3±0.17µs 44.7 MElem/sec 1.00 11.2±0.04µs 85.2 MElem/sec arrow_mono_points2/insert 1.01 1789.2±3.40ms 545.8 KElem/sec 1.00 1773.6±23.00ms 550.6 KElem/sec arrow_mono_points2/query 6.78 1102.4±18.79µs 885.9 KElem/sec 1.00 162.6±3.39µs 5.9 MElem/sec arrow_mono_strings2/insert 1.00 1777.3±5.89ms 549.5 KElem/sec 1.00 1777.3±7.53ms 549.5 KElem/sec arrow_mono_strings2/query 6.30 1149.9±15.36µs 849.3 KElem/sec 1.00 182.5±0.41µs 5.2 MElem/sec ``` --- Part of the primary caching series of PR (index search, joins, deserialization): - #4592 - #4593 - #4659 - #4680 - #4681 - #4698 - #4711 - #4712 - #4721 - #4726
teh-cmc
added a commit
that referenced
this pull request
Jan 10, 2024
Make it possible to toggle primary caching on and off at runtime, for both latest-at and range queries.  --- Part of the primary caching series of PR (index search, joins, deserialization): - #4592 - #4593 - #4659 - #4680 - #4681 - #4698 - #4711 - #4712 - #4721 - #4726
teh-cmc
added a commit
that referenced
this pull request
Jan 10, 2024
Integrates the cached APIs with the 2D & 3D spatial views, which is a pretty tough thing to do because there's a lot of abstraction going on in there. `main` vs. cache disabled vs. cache enable (5950X, Arch): ``` group main primcache_5_uncached primcache_5_cached ----- ---- -------------------- ------------------ Points3D/load_all 1.68 10.1±0.14ms 94.2 MElem/sec 1.00 6.0±0.07ms 157.9 MElem/sec 1.01 6.1±0.06ms 155.7 MElem/sec Points3D/load_colors 1.44 3.8±0.02ms 252.6 MElem/sec 1.00 2.6±0.05ms 364.0 MElem/sec 1.07 2.8±0.06ms 339.5 MElem/sec Points3D/load_picking_ids 15.16 1859.6±7.01µs 512.9 MElem/sec 1.01 124.3±3.92µs 7.5 GElem/sec 1.00 122.7±3.86µs 7.6 GElem/sec Points3D/load_positions 2.29 420.1±0.76µs 2.2 GElem/sec 1.03 189.3±7.44µs 4.9 GElem/sec 1.00 183.4±5.56µs 5.1 GElem/sec Points3D/load_radii 1.46 3.3±0.04ms 290.1 MElem/sec 1.05 2.4±0.03ms 404.8 MElem/sec 1.00 2.2±0.00ms 423.9 MElem/sec Points3D/query_archetype 2.51 676.1±7.59ns ? ?/sec 15859.98 4.3±0.06ms ? ?/sec 1.00 268.9±3.39ns ? ?/sec ``` --- Part of the primary caching series of PR (index search, joins, deserialization): - #4592 - #4593 - #4659 - #4680 - #4681 - #4698 - #4711 - #4712 - #4721 - #4726
teh-cmc
added a commit
that referenced
this pull request
Jan 10, 2024
Integrates the cached APIs with the TextLog & TimeSeries views, which is pretty trivial. This of course does nothing, since the cache doesn't cache range queries yet. --- Part of the primary caching series of PR (index search, joins, deserialization): - #4592 - #4593 - #4659 - #4680 - #4681 - #4698 - #4711 - #4712 - #4721 - #4726
teh-cmc
force-pushed
the
cmc/primcache_7_data_time
branch
from
88989df to
636eb92
Compare
teh-cmc
force-pushed
the
cmc/primcache_8_dedupe
branch
from
317c0ec to
a7e6e8a
Compare
4 tasks
4 tasks
teh-cmc
force-pushed
the
cmc/primcache_7_data_time
branch
2 times, most recently
from
b955acc to
6921781
Compare
teh-cmc
added a commit
that referenced
this pull request
Jan 15, 2024
) _99% grunt work, the only somewhat interesting thing happens in `query_archetype`_ Our query model always operates with two distinct timestamps: the timestamp you're querying for (`query_time`) vs. the timestamp of the data you get back (`data_time`). This is the result of our latest-at semantics: a query for a point at time `10` can return a point at time `2`. This is important to know when caching the data: a query at time `4` and a query at time `8` that both return the data at time `2` must share the same single entry or the memory budget would explode. This PR just updates all existing latest-at APIs so they return the data time in their response. This was already the case for range APIs. Note that in the case of `query_archetype`, which is a compound API that emits multiple queries, the data time of the final result is the most recent data time among all of its components. A follow-up PR will use the data time to deduplicate entries in the latest-at cache. --- Part of the primary caching series of PR (index search, joins, deserialization): - #4592 - #4593 - #4659 - #4680 - #4681 - #4698 - #4711 - #4712 - #4721 - #4726 - #4773 - #4784 - #4785 - #4793 - #4800
teh-cmc
force-pushed
the
cmc/primcache_8_dedupe
branch
from
a7e6e8a to
3252a8c
Compare
teh-cmc
force-pushed
the
cmc/primcache_8_dedupe
branch
from
1a41374 to
95fdb2e
Compare
teh-cmc
added a commit
that referenced
this pull request
Jan 15, 2024
Introduces a dedicated cache bucket for timeless data and properly forwards the information through all APIs downstream. --- Part of the primary caching series of PR (index search, joins, deserialization): - #4592 - #4593 - #4659 - #4680 - #4681 - #4698 - #4711 - #4712 - #4721 - #4726 - #4773 - #4784 - #4785 - #4793 - #4800
teh-cmc
added a commit
that referenced
this pull request
Jan 15, 2024
This implements cache invalidation via a `StoreSubscriber`. We keep track of the timestamps to invalidate in the `StoreSubscriber`, but we only do the actual removal of components at query time. This is similar to how we handle bucket sorting in the main store: doing it at query time has the benefit that the frame time effectively behaves as natural micro-batching mechanism that vastly improves performance. --- Part of the primary caching series of PR (index search, joins, deserialization): - #4592 - #4593 - #4659 - #4680 - #4681 - #4698 - #4711 - #4712 - #4721 - #4726 - #4773 - #4784 - #4785 - #4793 - #4800
teh-cmc
added a commit
that referenced
this pull request
Jan 15, 2024
) The primary cache now tracks memory statistics and display them in the memory panel. This immediately highlights a very stupid thing that the cache does: missing optional components that have been turned into streams of default values by the `ArchetypeView` are materialized as such :man_facepalming: - #4779 https://github.com/rerun-io/rerun/assets/2910679/876b264a-3f77-4d91-934e-aa8897bb32fe - Fixes #4730 --- Part of the primary caching series of PR (index search, joins, deserialization): - #4592 - #4593 - #4659 - #4680 - #4681 - #4698 - #4711 - #4712 - #4721 - #4726 - #4773 - #4784 - #4785 - #4793 - #4800
teh-cmc
added a commit
that referenced
this pull request
Jan 15, 2024
**Prefer on a per-commit basis, stuff has moved around** Range queries are back!... in the most primitive form possible. No invalidation, no bucketing, no optimization, no nothing. Just putting everything in place. https://github.com/rerun-io/rerun/assets/2910679/a65281e4-9843-4598-9547-ce7e45197995 --- Part of the primary caching series of PR (index search, joins, deserialization): - #4592 - #4593 - #4659 - #4680 - #4681 - #4698 - #4711 - #4712 - #4721 - #4726 - #4773 - #4784 - #4785 - #4793 - #4800
teh-cmc
added a commit
that referenced
this pull request
Jan 15, 2024
#4785) Title. https://github.com/rerun-io/rerun/assets/2910679/cf2c2748-a461-49fe-8124-c2a94164c956 --- Part of the primary caching series of PR (index search, joins, deserialization): - #4592 - #4593 - #4659 - #4680 - #4681 - #4698 - #4711 - #4712 - #4721 - #4726 - #4773 - #4784 - #4785 - #4793 - #4800
teh-cmc
added a commit
that referenced
this pull request
Jan 15, 2024
… range queries (#4793) Our low-level range APIs used to bake the latest-at results at `range.min - 1` into the range results, which is a big problem in a multi tenant setting because `range(1, 10)` vs. `latestat(1) + range(2, 10)` are two completely different things. Side-effect: a plot with a window of len 1 now behaves as expected: https://github.com/rerun-io/rerun/assets/2910679/957ac367-35a6-4bea-9f40-59d51c556639 --- Part of the primary caching series of PR (index search, joins, deserialization): - #4592 - #4593 - #4659 - #4680 - #4681 - #4698 - #4711 - #4712 - #4721 - #4726 - #4773 - #4784 - #4785 - #4793 - #4800
teh-cmc
added a commit
that referenced
this pull request
Jan 15, 2024
The most obvious and most important performance optimization when doing cached range queries: only upsert data at the edges of the bucket / ring-buffer. This works because our buckets (well, singular, at the moment) are always dense. - #4793  - #4800:  --- Part of the primary caching series of PR (index search, joins, deserialization): - #4592 - #4593 - #4659 - #4680 - #4681 - #4698 - #4711 - #4712 - #4721 - #4726 - #4773 - #4784 - #4785 - #4793 - #4800
This was referenced Jan 18, 2024
teh-cmc
added a commit
that referenced
this pull request
Jan 23, 2024
Range queries used to A) return the frame a T-1, B) accumulate state starting at T-1 and then C) yield frames starting at T. A) was a huge issue for many reasons, which #4793 took care of by eliminating both A) and B). But we need B) for range queries to be context-free, i.e. to be guaranteed that `Range(5, 10)` and `Range(4, 10)` will return the exact same data for frame `5`. This is crucial for multi-tenant settings where those 2 example queries would share the same cache. It also is the nicer-nicer version of the range semantics that we wanted anyway, I just didn't realize back then that it would require so little changes, or I would've gone straight for that. --- Part of the primary caching series of PR (index search, joins, deserialization): - #4592 - #4593 - #4659 - #4680 - #4681 - #4698 - #4711 - #4712 - #4721 - #4726 - #4773 - #4784 - #4785 - #4793 - #4800 - #4851 - #4852 - #4853 - #4856
teh-cmc
added a commit
that referenced
this pull request
Jan 23, 2024
Simply add a timeless path for the range cache, and actually only iterate over the range the user asked for (we were still blindly iterating over everything until now). Also some very minimal clean up related to #4832, but we have a long way to go... - #4832 --- - Fixes #4821 --- Part of the primary caching series of PR (index search, joins, deserialization): - #4592 - #4593 - #4659 - #4680 - #4681 - #4698 - #4711 - #4712 - #4721 - #4726 - #4773 - #4784 - #4785 - #4793 - #4800 - #4851 - #4852 - #4853 - #4856
teh-cmc
added a commit
that referenced
this pull request
Jan 23, 2024
Implement range invalidation and do a quality pass over all the size tracking stuff in the cache. **Range caching is now enabled by default!** - Fixes #4809 - Fixes #374 --- Part of the primary caching series of PR (index search, joins, deserialization): - #4592 - #4593 - #4659 - #4680 - #4681 - #4698 - #4711 - #4712 - #4721 - #4726 - #4773 - #4784 - #4785 - #4793 - #4800 - #4851 - #4852 - #4853 - #4856
teh-cmc
added a commit
that referenced
this pull request
Jan 23, 2024
- Quick sanity pass over all the intermediary locks and refcounts to make sure we don't hold anything for longer than we need. - Get rid of all static globals and let the caches live with their associated stores in `EntityDb`. - `CacheKey` no longer requires a `StoreId`. --- - Fixes #4815 --- Part of the primary caching series of PR (index search, joins, deserialization): - #4592 - #4593 - #4659 - #4680 - #4681 - #4698 - #4711 - #4712 - #4721 - #4726 - #4773 - #4784 - #4785 - #4793 - #4800 - #4851 - #4852 - #4853 - #4856
teh-cmc
added
include in changelog
and removed
exclude from changelog
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Introduces the notion of cache deduplication: given a query at time
4and a query at time8that both returns data at time2, they must share a single cache entry.I.e. starting with this PR, scrubbing through the OPF example will not result if more cache memory being used.
Part of the primary caching series of PR (index search, joins, deserialization):
VecDequeextensions #4592FlatVecDeque#4593LatestAt(T-1)results into low-level range queries #4793Checklist
mainbuild: app.rerun.ionightlybuild: app.rerun.io