4 Commits (main)

Author	SHA1	Message	Date
Robert Fratto	f6091a67d1	chore(engine): move to toggleable section prefetching (#19142) ... dataset.readerDownloader was originally introduced in #16429, an attempt to balance peak memory usage of reading a section with read times by downloading a configurable size of pages in advance. In practice, each roundtrip to object storage adds too much of a latency hit, and we've started to set the cache limit high enough to ensure that each reader only needs a single prefetch. Given what we've found, it no longer makes sense to control peak memory usage via the prefetch size. Other options, such as downloading directly to disk, may be explored in the future. In the meantime, this PR removes the ability to specify a cache size. All non-pruned pages will be bulk requested using the range reader (#19067) on the first read call. Pages which have left the potential read window will continue to be eagerly removed for garbage collection. However, we don't want to prefetch when the dataset is entirely in memory, which is the case when the logs section builder is performing k-way merge over in-memory sections. To lower the memory usage of builders, prefetching is configurable. For this initial PR, prefetching is only disabled for the logs section builder; all other reads force prefetching. Signed-off-by: Robert Fratto <robertfratto@gmail.com>	4 months ago
Robert Fratto	595715146f	chore(dataobj)!: introduce new centralized dataset encoding format (#18871) ... Signed-off-by: Robert Fratto <robertfratto@gmail.com>	4 months ago
Ashwanth	da4e59f8c6	chore(dataobj): store sort order information in section metadata (#18499)	5 months ago
Periklis Tsirakidis	6eedcaaded	chore(dataobj): Add new index pointers section type (#18441)	6 months ago
Robert Fratto	06cf52757b	chore(dataobj): Add columnar reading APIs to logs and streams sections (#17976) ... This PR introduces columnar reading APIs to both the logs and streams sections. As part of this change, both sections now expose the columns stored in the section, which are then used to define predicates on the readers. The columnar reading APIs emit a sequence of `arrow.Record`s, where each record has no more than the batch size passed to the `Reader.Read` method. Unlike the original row-based readers, the columnar readers: * Allow reading a subset of columns * Have column-based predicates that map literally to the new query engine To be able to represent a column's value for use in predicates, we import [github.com/apache/arrow-go/v18/arrow/scalar](https://pkg.go.dev/github.com/apache/arrow-go/v18@v18.3.0/arrow/scalar). Section Reader implementations perform mapping to and from the internal `dataset.Value`. The implementation of `logs.Reader` is almost an identical copy of the `streams.Reader` implementation. I've opted for duplicating the implementation in the short term, since there's no obvious way for how to deduplicate it in a way keeps the package structure easy to understand. We can change this in the future if a clean and simple approach presents itself. This PR does not yet update `DataObjScan` to make use of these new APIs; that's being left out of scope for another PR to handle.	7 months ago
Robert Fratto	36bcb5aa2d	chore(dataobj): fix regression in allocations of row readers (#17784)	7 months ago
Robert Fratto	a237444f0d	refactor(dataobj): invert dependency between dataobj and sections (#17762) ... Originally, the dataobj package was a higher-level API around sections. This design caused it to become a bottleneck: * Implementing any new public behaviour for a section required bubbling it up to the dataobj API for it to be exposed, making it tedious to add new sections or update existing ones. * The `dataobj.Builder` pattern was focused on constructing dataobjs for storing log data, which will cause friction as we build objects around other use cases. This PR builds on top of the foundation laid out by #17704 and #17708, fully inverting the dependency between dataobj and sections: * The `dataobj` package has no knowledge of what sections exist, and can now be used for writing and reading generic sections. Section packages now create higher-level APIs around the abstractions provided by `dataobj`. * Section packages are now public, and callers interact directly with these packages for writing and reading section-specific data. * All logic for a section (encoding, decoding, buffering, reading) is now fully self-contained inside the section package. Previously, the implementation of each section was spread across three packages (`pkg/dataobj/internal/encoding`, `pkg/dataobj/internal/sections/SECTION`, `pkg/dataobj`). * Cutting a section is now a decision made by the caller rather than the section implementation. Previously, the logs section builder would create multiple sections. For the most part, this change is a no-op, with two exceptions: 1. Section cutting is now performed by the caller; however, this shouldn't result in any issues. 2. Removing the high-level `dataobj.Stream` and `dataobj.Record` types will temporarily reduce the allocation gains from #16988. I will address this after this PR is merged.	7 months ago
Robert Fratto	7b0d6d7144	chore(dataobj): Migrate to dynamic section types (#17708)	8 months ago
Robert Fratto	3fbebd3bc4	chore(dataobj): migrate to section-relative offsets (#17704)	8 months ago
benclive	e5784d7a82	perf: Reuse Values in dataobj.Reader to reduce allocs (#16988) ... Co-authored-by: Robert Fratto <robertfratto@gmail.com>	9 months ago
Robert Fratto	bd4331363b	chore: align benchmark results between chunks and dataobjs (#17127)	9 months ago
Robert Fratto	9be95a3e8c	chore(dataobj): reduce memory overhead of merging logs stripes (#16833)	9 months ago
Robert Fratto	cd11367148	chore(dataobj): Wire dataset.Reader into dataobj.LogsReader and dataobj.StreamsReader (#16701) ... Co-authored-by: Ashwanth Goli <iamashwanth@gmail.com>	10 months ago
Robert Fratto	7f0dc01e21	chore(dataobj): introduce more efficient dataset reading (#16429) ... This commit introduces a new API, dataset.Reader, which scans through values in datasets more efficiently than dataset.Iter. As a result, dataset.Iter has been removed in favour of the new Reader. dataset.Reader is more efficient than dataset.Iter because: - It avoids using iter.Pull; the number of values flowing through coroutines added a significant amount of CPU overhead. - It supports efficient uses of predicates, where non-predicate columns are only loaded after finding matching rows from predicate columns; this avoids loading unnecessary pages. - It adds initial support for bulk page downloading, so that we need fewer roundtrips to object storage when iterating over a dataset. (This is not fully implemented yet; see below). - Early performance benchmarks see a ~3.4x speed improvement over reading 7GB of uncompressed logs, going from 192 MB/s (with dataset.Iter) to 652 MB/s (with dataset.Reader). This commit is truly massive, and to avoid it getting any larger, I'm leaving some follow up work to additional commits: - Updating dataobj.LogsReader and dataobj.StreamsReader to use the new API. - Adding support for downloading pages in batches; dataset.Reader will request multiple pages at once, but our decoder still unwraps that into several GET requests to object storage. Signed-off-by: Robert Fratto <robertfratto@gmail.com> Co-authored-by: Cyril Tovena <cyril.tovena@gmail.com>	10 months ago
Ashwanth	651d410eea	chore(dataobj): add "uncompressed size" column to streams section (#16322)	10 months ago
Robert Fratto	1adb1e39d3	chore(dataobj): add initial high-level APIs for reading streams and log records (#15974)	11 months ago
Robert Fratto	11714a71c7	chore(dataobj): add logs tracking and encoding (#15757)	12 months ago
Robert Fratto	d1fd0b9ffc	chore(dataobj): add streams tracking and encoding (#15713)	12 months ago