- treat OOM protection as "resource exhausted"
- use `DataFusionError` in more places instead of opaque `Box<dyn Error>`
- improve conversion from/into `DataFusionError` to preserve more
semantics
Overall, this improves our error handling. DF can now return errors like
"resource exhausted" and gRPC should now automatically generate a
sensible status code for it.
Fixes#5799.
* feat: send only needed projection columns from querier to ingester in case of normal SQL queries
* refactor: push column index down until we need to convert them strings
* fix: make the test deterministic
* test: test for the projection pushdown
* test: add asserts for the proj pushdown test
* test: implement projection pushdown for partitions of MockIngesterConnection
* chore: cleanup
* chore: address review comments
* chore: Apply suggestions from code review
Co-authored-by: Andrew Lamb <andrew@nerdnetworks.org>
* refactor: address review comments
Co-authored-by: Andrew Lamb <andrew@nerdnetworks.org>
Co-authored-by: kodiakhq[bot] <49736102+kodiakhq[bot]@users.noreply.github.com>
* chore: Upgrade to Rust 1.64
* fix: Use iter find instead of a for loop, thanks clippy
* fix: Remove some needless borrows, thanks clippy
* fix: Use then_some rather than then with a closure, thanks clippy
* fix: Use iter retain rather than filter collect, thanks clippy
Co-authored-by: kodiakhq[bot] <49736102+kodiakhq[bot]@users.noreply.github.com>
* refactor: do not run de-dup in ingester for querier requests
This removes the entire de-dup logic from the inegster for querier
requests. Furthermore, it even removes the entire datafusion execution
from the querier and just dumps the in-memory record batches as quickly
as possible. No filters are applied. Note that even prior to this PR,
we've never applied projections (tracked by #5624).
**Pros:**
- speed up query planning within the querier (since we need the ingester
response for state reconciling)
- lowered ingester CPU load
**Cons:**
- more querier<>ingester network traffic
Closes#5602.
* test: extend query test case
* fix: ingester tests
Co-authored-by: kodiakhq[bot] <49736102+kodiakhq[bot]@users.noreply.github.com>
* refactor: arc the cached table
* refactor: use cheaper hash keys for projected schemas
Instead of using the column names to address projected schemas, let's
use the column IDs.
Co-authored-by: kodiakhq[bot] <49736102+kodiakhq[bot]@users.noreply.github.com>
* fix: retry ingester requests faster
The retries introduced in #5695 are too slow and block the entire
querier for minutes (until the very long gRPC timeout kicks in).
* fix: add error details on why the query planning failed
Co-authored-by: kodiakhq[bot] <49736102+kodiakhq[bot]@users.noreply.github.com>
* refactor: retry querier->ingester requests
Esp. for InfluxRPC requests that scan multiple tables, it may be that
one ingester requests fails. We shall retry that request instead of
failing the entire query.
* refactor: improve docs
Co-authored-by: Andrew Lamb <andrew@nerdnetworks.org>
* fix: less foo
* docs: remove outdated TODO
* test: assert that panic happened
Co-authored-by: Andrew Lamb <andrew@nerdnetworks.org>
* refactor: improve consistent access under "remove if"
With all the concurrency introduced in #5668, we should be a bit more
careful with our "remove if" handling, esp. if a removal is triggered
while a load is running concurrently. This change introduces as
`remove_if_and_get` helper that ensures this and the querier over to use
it. The parquet file and tombstone caches required a bit of a larger
change because there the invalidation and the actual GET were kinda
separate. We had this separation for the other caches as well at some
point and decided that this easily leads to API misuse, so I took this
opportunity to "fix" the parquet file and tombstone cache as well.
* docs: improve
* feat: split "pruned" metric into "early" and "late"
* docs: improve
Co-authored-by: Andrew Lamb <andrew@nerdnetworks.org>
* docs: explain `PruningMetrics`
* test: try to test pruning
Co-authored-by: Andrew Lamb <andrew@nerdnetworks.org>
Create chunks in querier concurrently after we've pre-filtered them.
Chunk creation still may require a bit of cached information (e.g. the
partition sort key) and we can easily fetch these concurrently instead
of in order.
Co-authored-by: kodiakhq[bot] <49736102+kodiakhq[bot]@users.noreply.github.com>
This should lower catalog load and eliminate a few costly cache misses.
Co-authored-by: kodiakhq[bot] <49736102+kodiakhq[bot]@users.noreply.github.com>
In our data model, a chunk always belongs to a partition[^1], so let's
not make this attribute optional. The optional value only leads to
-- mostly surprising -- conditional behavior, ranging from "do not equalize
the partition sort key" (querier) to "always consider the chunk overlapping"
(iox_query when dealing with ingester chunks).
[^1]: This is even true when the chunk belongs to a parquet file that is not
yet added to the catalog, contrary to what a comment in the ingester
stated. The catalog and data model used by the querier are two totally
different things.
* refactor: read querier parquet files from cache
* refactor: only use parquet files in querier (no RB)
Co-authored-by: kodiakhq[bot] <49736102+kodiakhq[bot]@users.noreply.github.com>
* ci: use same feature set in `build_dev` and `build_release`
* ci: also enable unstable tokio for `build_dev`
* chore: update tokio to 1.21 (to fix console-subscriber 0.1.8
* fix: "must use"
Remove our own hand-rolled logic and let DataFusion read the parquet
files.
As a bonus, this now supports predicate pushdown to the deserialization
step, so we can use parquets as in in-mem buffer.
Note that this currently uses some "nested" DataFusion hack due to the
way the `QueryChunk` interface works. Midterm I'll change the interface
so that the `ParquetExec` nodes are directly visible to DataFusion
instead of some opaque `SendableRecordBatchStream`.
* refactor: do not override parquet file size in querier
This is going to be an issue when we actually rely on the size for
reading, see #5531.
* refactor: use selected file size mocking in compactor
Do not blindly override parquet file sizes for all subsystems.
This is going to be an issue when we actually rely on the size for
reading, see #5531.
* refactor: remove ability to override file sizes in catalog
Blindly overriding data for all subsystems is dangerous, because some
parts of our stack actually rely on the actual file size. See #5531.
* docs: explain `size_overrides`
The API user may still use a `Box<dyn ...>` if they want, but they
technically don't have to.
Co-authored-by: kodiakhq[bot] <49736102+kodiakhq[bot]@users.noreply.github.com>
The API user still CAN use dynamic dispatch but doesn't have to. This
also simplifies the generics a bit.
This is similar to #5520.
Co-authored-by: kodiakhq[bot] <49736102+kodiakhq[bot]@users.noreply.github.com>
This removes some `Box<dyn ...>` indirection when the user doesn't want
it (you still can, but don't have to) and makes the whole type handling
easier to understand.
Co-authored-by: kodiakhq[bot] <49736102+kodiakhq[bot]@users.noreply.github.com>
1. Cache converted schema instead of catalog schema. This safes a buch
of memcopies during conversion.
2. Simplify creation of new chunks, we now only need a `CachedTable`
instead of a namespace and a table schema.
In an artificial benchmark, this removed around 10ms from the query
(although that was prior to #5467 which moved schema conversion one
level up). Still I think it is the cleaner cache design.
Co-authored-by: kodiakhq[bot] <49736102+kodiakhq[bot]@users.noreply.github.com>
* refactor: use a single timestamp in policy backend
Prior to this PR we had at least 1 `TimeProvider::now` calls per GET
request (for caches that only used LRU) and up to 3 calls (caches with
LRU + refresh + TTL). Let's instead use a single timestamp that is
created by the policy backend itself (instead of the policies). This has
the following consequences:
- **efficiency:** `SystemProvider::now` is not free, even though under Linux
this doesn't result in a syscall, it uses the stdlib time system which
also checks for monotonicity
- **consistency:** All changes for a single trigger (e.g. a
GET cache call) now use a single timestamp instead of slightly
increasing ones. I argue this is the better semantic, simpler to
understand and better to debug.
For some (slightly artificial) local performance experiment, this shaves
off around 2ms per single-table SQL query. However I expect that there might
be more degenerated cases (e.g. multi-table SQL queries or some
InfluxRPC requests that hit multiple tables).
The majority of this patch is moving the `TimeProvider` from the
policies into the policy backend.
* docs: explain `now` parameter
* fix: hoist repeated computation out of chunk creation
We have hundreds of chunks per table, so it is beneficial to only
do common work once.
* chore: remove TableCache as it is no longer used
* fix: prune chunks both before and after metadata fetch
Fetching the metadata for all the chunks in a table is expensive,
especially when we have a narrow time range query that only
needs a few chunks.
* chore: fix clippy
* fix: fix up some last tests
* fix: review comments
Co-authored-by: kodiakhq[bot] <49736102+kodiakhq[bot]@users.noreply.github.com>
This doesn't really need to be fallible but forces propagation of a ton
of error handling - no shards is always a sign of something being very
wrong, and can be caught in the caller if it's for some reason an
acceptable state / can be recovered from.
* refactor: allow `ChangeRequest` to carry a lifetime
Let's not restrict our change functions to `'static` because this would
require us to clone loads of data to achieve predicate-based
`remove_if`.
* refactor: convert `remove_if` feature to policy framework
Decided to drop the "shared" functionality. We only use the small
`remove_if` bit which is way easier to reason about.
For #5320.
* refactor: address review comments
Co-authored-by: kodiakhq[bot] <49736102+kodiakhq[bot]@users.noreply.github.com>
* refactor: port TTL backend to policy framework
Note that this is "just" a port, it does NOT change how TTL works. This
will be done in #5318.
Helps with #5320.
* fix: ensure inner backend is empty
* test: add some smoke test
We already prune all chunks in the query-access layer. There's no need
to do that another time (which is actually the first time) in
`QuerierTable::chunks`. The time savings we get from feeding less chunks
into the state reconciling should be negligible. On the pro-side however
we get a more streamlined data flow and actually correct chunk pruning
metrics. Also see #5336.
Co-authored-by: kodiakhq[bot] <49736102+kodiakhq[bot]@users.noreply.github.com>
- emit a warning if we cannot even attempt to prune chunks due to an
error. This is always either a missing feature or a bug (even though
it does not impact correctness but _only_ performance). Also see
https://github.com/influxdata/conductor/issues/1107
- change metrics to clearly differentiate between "could not prune" and
"not pruned"
- add new "not pruned" observer hook (this was missing for some reason,
the "pruned" hook existed though)
* refactor: make could-not-prune reason a static string
* refactor: introduce `QuerierTableArgs`
* feat: chunk pruning metrics
Closes#4974.
* refactor: address review comments
* refactor: use static typing for not-pruned reason
* refactor: pass chunk to not-pruned observer and use it for some metrics
Co-authored-by: kodiakhq[bot] <49736102+kodiakhq[bot]@users.noreply.github.com>
* feat: make querier RAM pool split a proper feature
- use propre pool names
- expose sizing via CLI/env
Closes https://github.com/influxdata/conductor/issues/1102.
* refactor: improve naming and docs
Co-authored-by: Andrew Lamb <andrew@nerdnetworks.org>
Co-authored-by: Andrew Lamb <andrew@nerdnetworks.org>
Co-authored-by: kodiakhq[bot] <49736102+kodiakhq[bot]@users.noreply.github.com>
Quick&Dirty implementation of a RAM-pool split to see if this has any
effect. I expect the querier performance to improve due to this because
large read buffers can no longer evict precious metadata.
Co-authored-by: kodiakhq[bot] <49736102+kodiakhq[bot]@users.noreply.github.com>
This is what DataFusion uses by default and I don't see a reason why we
should use such small batch sizes.
The affect is probably only visible in certain filter-aggregate queries
that don't focus on a single series (because there we likely end up with
1 or 2 batches only, esp. after #5250) for coarse-grained filters, esp.
when the filter key is not the first sort key.
Co-authored-by: kodiakhq[bot] <49736102+kodiakhq[bot]@users.noreply.github.com>
* feat: `QueryChunk::as_any`
* feat: allo `ChunkPruner::prune_chunks` to fail
* feat: limit per-table chunk data for every query
Closes#5211.
* fix: address review comments
Co-authored-by: Andrew Lamb <andrew@nerdnetworks.org>
Co-authored-by: Andrew Lamb <andrew@nerdnetworks.org>
* refactor: remove min_sequnce_number
* fix: typos
* fix: remove min_sequencer_number from new files from merging main
* fix: add back throwing error if the compactor compacts files persisted by the ingester after the ingester sends max seq_num back to querier
* test: add test_compactor_collision back but modify the input to make it work woth new changes
Co-authored-by: kodiakhq[bot] <49736102+kodiakhq[bot]@users.noreply.github.com>
* feat: cache tracing
Add tracing to the metrics cache wrapper. The extra arguments for GET
and PEEK make this quite simple, because the wrapper can just extend the
inner args with the trace information.
We currently terminate the span in `querier::cache` (i.e. only pass in
`None`, so no tracing will occur) to keep this PR rather small. This
will be changed in subsequent PRs.
For #5129.
* fix: typo
Co-authored-by: Andrew Lamb <andrew@nerdnetworks.org>
Co-authored-by: Andrew Lamb <andrew@nerdnetworks.org>
Co-authored-by: kodiakhq[bot] <49736102+kodiakhq[bot]@users.noreply.github.com>
- remove `IOxSessionContext::default()` because untracked contexts
should only be created by tests
- remove `Option<IOxSessionContext>` because it is a typed workaround
for `IOxSessionContext::default`
Tests should use `IOxSessionContext::testing` and all _normal_ users
should create proper contexts.
I suspect this will help tracing or at least prevent silent regressions.
See #5129.
Co-authored-by: kodiakhq[bot] <49736102+kodiakhq[bot]@users.noreply.github.com>
This will be used to pass spans down to `CacheWithMetrics` (or a new
wrapper specific to tracing) and will help with #5129.
Co-authored-by: kodiakhq[bot] <49736102+kodiakhq[bot]@users.noreply.github.com>
This adds tracing of querire->ingester request up to the point where we
perform the network request, i.e. the trace will only appear on the
querier side. We may extend this at some point to carry the tracing
information to the ingester as well.
Ref #5129.
Co-authored-by: kodiakhq[bot] <49736102+kodiakhq[bot]@users.noreply.github.com>
I forgot to address a TODO in #5091. Extends to test to actually check
the chunk stage and removes the function for manual force-loads.
Co-authored-by: kodiakhq[bot] <49736102+kodiakhq[bot]@users.noreply.github.com>
* refactor: remove parquet chunk ID to `ChunkMeta`
* refactor: return `Arc` from `QueryChunk::summary`
This is similar to how we handle other chunk data like schemas. This
allows a chunk to change/refine its "believe" over its own payload while
it is passed around in the query stack.
Helps w/ #5032.
This is not relevant at the moment for prod since other layers
prevent/filter queries for non-existing namespaces.
However this messes up the flux integration tests, see
https://github.com/influxdata/conductor/issues/997
So let's disable this specific cache case until #4617 is implemented
which may be used by the flux tests.
Fixes https://github.com/influxdata/conductor/issues/997
Co-authored-by: kodiakhq[bot] <49736102+kodiakhq[bot]@users.noreply.github.com>
Instead of using some hand-rolled timestamp-based logic (or just
"unknown") all over the place, just use logic introduced in #5017.
This requires slightly improved table summaries within the querier that
at least has min/max for the timestamp column. For that, the former
`IngesterChunk`-specific `calculate_summary` method was extended to
`create_basic_summary` to include that data and is now also used by
`QuerierParquetChunk`.
Note: `QuerierRBChunk` already has detailled metrics that are provided
by the read buffer implementation.
Should we ever need even better pruning for `QuerierParquetChunk` (or
`IngesterChunk`) then we _only_ need add extra data to the table
summaries.
Closes#4976.
Co-authored-by: kodiakhq[bot] <49736102+kodiakhq[bot]@users.noreply.github.com>
Marco Neumann
Dom Dwyer
Andrew Lamb
Nga Tran
Carol (Nichols || Goulding)
dependabot[bot]
YIXIAO SHI
Luke Bond
Carol (Nichols || Goulding)
Jake Goulding
Sam Arnold
kodiakhq[bot]