CNDB-13639: Adds optional prefetching for known sequential reads#1692
Merged
CNDB-13639: Adds optional prefetching for known sequential reads#1692
Conversation
Checklist before you submit for review
|
pcmanus
force-pushed
the
CNDB-13639-prefetching
branch
from
3ca9092 to
18c9e57
Compare
blambov
approved these changes
Apr 14, 2025
blambov
left a comment
blambov
left a comment
There was a problem hiding this comment.
Looks good to me.
Pushed a patch with a couple of typo corrections.
This introduces a `PrefetchingRebufferer` that, when enabled, prefetches (read from the underlying reader) a configurable number of chunks in advance (so only work on top of rebufferer factories that work with chunks, i.e. not uncompressed mmap). Prefetching must first be enabled globally by setting `-Dcassandra.read_prefetching_size_kb` to the desired value. With that, and assuming the disk access mode allows it (meaning, it's not uncompressed mmap), then prefetching will be applied to reads that are "clearly" sequential. Mostly, as of this patch, this means the sstable "scanners" and `SortedStringTableCursor`, so compaction, SAI index building and tools that read sstable fully (scrub, verifier) will benefit from it. Since rebufferers are synchronous, prefetching is done through a fixed thread pool. The number of thread of that pool can be set with `-Dcassandra.read_prefetching_threads` (but default to the number of "processors"). The `-Dcassandra.read_prefetching_window` can also be set to define how often prefetching is re-triggered. By default, when there is less than half of the value of `-Dcassandra.read_prefetching_size_kb` prefetched, then prefetching is triggered.
pcmanus
force-pushed
the
CNDB-13639-prefetching
branch
from
a14e75e to
070433b
Compare
|
❌ Build ds-cassandra-pr-gate/PR-1692 rejected by Butler1 new test failure(s) in 4 builds Found 1 new test failures
Found 6 known test failures |
djatnieks
pushed a commit
that referenced
this pull request
May 13, 2025
This introduces a `PrefetchingRebufferer` that, when enabled, prefetches (read from the underlying reader) a configurable number of chunks in advance (so only work on top of rebufferer factories that work with chunks, i.e. not uncompressed mmap). Prefetching must first be enabled globally by setting `-Dcassandra.read_prefetching_size_kb` to the desired value. With that, and assuming the disk access mode allows it (meaning, it's not uncompressed mmap), then prefetching will be applied to reads that are "clearly" sequential. Mostly, as of this patch, this means the sstable "scanners" and `SortedStringTableCursor`, so compaction, SAI index building and tools that read sstable fully (scrub, verifier) will benefit from it. Since rebufferers are synchronous, prefetching is done through a fixed thread pool. The number of thread of that pool can be set with `-Dcassandra.read_prefetching_threads` (but default to the number of "processors"). The `-Dcassandra.read_prefetching_window` can also be set to define how often prefetching is re-triggered. By default, when there is less than half of the value of `-Dcassandra.read_prefetching_size_kb` prefetched, then prefetching is triggered. See riptano/cndb#13639 for additional motivation. I'll note that this patch is adapted from the similar DSE behavior. However, there is a fair bit of modification since the DSE version was relying on the asynchronous nature of rebufferers there (it also had a few behavior I didn't fully understood and I didn't dig when it felt a bit specific to DSE). One point worth mentioning was that the DSE version was relying on the ability of the underlying channel to create batches when multiple chunks are prefetched. This is something that could have advantages and we could try to add back over time, but it's a lot harder to see how to do that in the non-async world of C*. Typically, in DSE, since all the prefetches were initiated on the current thread (without blocking it), building the "batch" was relatively easy, but it doesn't translate in this patch. Not to mention that currently the underlying channels have no batching options and those would have to be added (in a way that trickled down to the channel implementation within CNDB). Anyway, as of this patch, chunks are prefetched in parallel but with no batching optimization. Which does mean the `window' parameter is probably not as useful as for DSE, but I've kept it nonetheless as it's not exactly adding complexity. In the context of compaction in CNDB though, I think this mean that it would kind of be ideal if we could use a relatively large "chunk size" for the readers: we read the full file anyway, so there is no waste there to use a large-ish chunk size, and it would kind of provide batching for prefetching "for free" (in the sense that if we want to prefetch, say, 128kb in advance, then we only need 2 chunks with a 64kb chunk, but need a lot more with a 4kb chunk size). --------- Co-authored-by: Branimir Lambov <branimir.lambov@datastax.com>
djatnieks
pushed a commit
that referenced
this pull request
May 18, 2025
This introduces a `PrefetchingRebufferer` that, when enabled, prefetches (read from the underlying reader) a configurable number of chunks in advance (so only work on top of rebufferer factories that work with chunks, i.e. not uncompressed mmap). Prefetching must first be enabled globally by setting `-Dcassandra.read_prefetching_size_kb` to the desired value. With that, and assuming the disk access mode allows it (meaning, it's not uncompressed mmap), then prefetching will be applied to reads that are "clearly" sequential. Mostly, as of this patch, this means the sstable "scanners" and `SortedStringTableCursor`, so compaction, SAI index building and tools that read sstable fully (scrub, verifier) will benefit from it. Since rebufferers are synchronous, prefetching is done through a fixed thread pool. The number of thread of that pool can be set with `-Dcassandra.read_prefetching_threads` (but default to the number of "processors"). The `-Dcassandra.read_prefetching_window` can also be set to define how often prefetching is re-triggered. By default, when there is less than half of the value of `-Dcassandra.read_prefetching_size_kb` prefetched, then prefetching is triggered. See riptano/cndb#13639 for additional motivation. I'll note that this patch is adapted from the similar DSE behavior. However, there is a fair bit of modification since the DSE version was relying on the asynchronous nature of rebufferers there (it also had a few behavior I didn't fully understood and I didn't dig when it felt a bit specific to DSE). One point worth mentioning was that the DSE version was relying on the ability of the underlying channel to create batches when multiple chunks are prefetched. This is something that could have advantages and we could try to add back over time, but it's a lot harder to see how to do that in the non-async world of C*. Typically, in DSE, since all the prefetches were initiated on the current thread (without blocking it), building the "batch" was relatively easy, but it doesn't translate in this patch. Not to mention that currently the underlying channels have no batching options and those would have to be added (in a way that trickled down to the channel implementation within CNDB). Anyway, as of this patch, chunks are prefetched in parallel but with no batching optimization. Which does mean the `window' parameter is probably not as useful as for DSE, but I've kept it nonetheless as it's not exactly adding complexity. In the context of compaction in CNDB though, I think this mean that it would kind of be ideal if we could use a relatively large "chunk size" for the readers: we read the full file anyway, so there is no waste there to use a large-ish chunk size, and it would kind of provide batching for prefetching "for free" (in the sense that if we want to prefetch, say, 128kb in advance, then we only need 2 chunks with a 64kb chunk, but need a lot more with a 4kb chunk size). --------- Co-authored-by: Branimir Lambov <branimir.lambov@datastax.com>
michaelsembwever
pushed a commit
that referenced
this pull request
Feb 6, 2026
This introduces a `PrefetchingRebufferer` that, when enabled, prefetches (read from the underlying reader) a configurable number of chunks in advance (so only work on top of rebufferer factories that work with chunks, i.e. not uncompressed mmap). Prefetching must first be enabled globally by setting `-Dcassandra.read_prefetching_size_kb` to the desired value. With that, and assuming the disk access mode allows it (meaning, it's not uncompressed mmap), then prefetching will be applied to reads that are "clearly" sequential. Mostly, as of this patch, this means the sstable "scanners" and `SortedStringTableCursor`, so compaction, SAI index building and tools that read sstable fully (scrub, verifier) will benefit from it. Since rebufferers are synchronous, prefetching is done through a fixed thread pool. The number of thread of that pool can be set with `-Dcassandra.read_prefetching_threads` (but default to the number of "processors"). The `-Dcassandra.read_prefetching_window` can also be set to define how often prefetching is re-triggered. By default, when there is less than half of the value of `-Dcassandra.read_prefetching_size_kb` prefetched, then prefetching is triggered. See riptano/cndb#13639 for additional motivation. I'll note that this patch is adapted from the similar DSE behavior. However, there is a fair bit of modification since the DSE version was relying on the asynchronous nature of rebufferers there (it also had a few behavior I didn't fully understood and I didn't dig when it felt a bit specific to DSE). One point worth mentioning was that the DSE version was relying on the ability of the underlying channel to create batches when multiple chunks are prefetched. This is something that could have advantages and we could try to add back over time, but it's a lot harder to see how to do that in the non-async world of C*. Typically, in DSE, since all the prefetches were initiated on the current thread (without blocking it), building the "batch" was relatively easy, but it doesn't translate in this patch. Not to mention that currently the underlying channels have no batching options and those would have to be added (in a way that trickled down to the channel implementation within CNDB). Anyway, as of this patch, chunks are prefetched in parallel but with no batching optimization. Which does mean the `window' parameter is probably not as useful as for DSE, but I've kept it nonetheless as it's not exactly adding complexity. In the context of compaction in CNDB though, I think this mean that it would kind of be ideal if we could use a relatively large "chunk size" for the readers: we read the full file anyway, so there is no waste there to use a large-ish chunk size, and it would kind of provide batching for prefetching "for free" (in the sense that if we want to prefetch, say, 128kb in advance, then we only need 2 chunks with a 64kb chunk, but need a lot more with a 4kb chunk size). --------- Co-authored-by: Branimir Lambov <branimir.lambov@datastax.com>
michaelsembwever
pushed a commit
that referenced
this pull request
Feb 10, 2026
This introduces a `PrefetchingRebufferer` that, when enabled, prefetches (read from the underlying reader) a configurable number of chunks in advance (so only work on top of rebufferer factories that work with chunks, i.e. not uncompressed mmap). Prefetching must first be enabled globally by setting `-Dcassandra.read_prefetching_size_kb` to the desired value. With that, and assuming the disk access mode allows it (meaning, it's not uncompressed mmap), then prefetching will be applied to reads that are "clearly" sequential. Mostly, as of this patch, this means the sstable "scanners" and `SortedStringTableCursor`, so compaction, SAI index building and tools that read sstable fully (scrub, verifier) will benefit from it. Since rebufferers are synchronous, prefetching is done through a fixed thread pool. The number of thread of that pool can be set with `-Dcassandra.read_prefetching_threads` (but default to the number of "processors"). The `-Dcassandra.read_prefetching_window` can also be set to define how often prefetching is re-triggered. By default, when there is less than half of the value of `-Dcassandra.read_prefetching_size_kb` prefetched, then prefetching is triggered. See riptano/cndb#13639 for additional motivation. I'll note that this patch is adapted from the similar DSE behavior. However, there is a fair bit of modification since the DSE version was relying on the asynchronous nature of rebufferers there (it also had a few behavior I didn't fully understood and I didn't dig when it felt a bit specific to DSE). One point worth mentioning was that the DSE version was relying on the ability of the underlying channel to create batches when multiple chunks are prefetched. This is something that could have advantages and we could try to add back over time, but it's a lot harder to see how to do that in the non-async world of C*. Typically, in DSE, since all the prefetches were initiated on the current thread (without blocking it), building the "batch" was relatively easy, but it doesn't translate in this patch. Not to mention that currently the underlying channels have no batching options and those would have to be added (in a way that trickled down to the channel implementation within CNDB). Anyway, as of this patch, chunks are prefetched in parallel but with no batching optimization. Which does mean the `window' parameter is probably not as useful as for DSE, but I've kept it nonetheless as it's not exactly adding complexity. In the context of compaction in CNDB though, I think this mean that it would kind of be ideal if we could use a relatively large "chunk size" for the readers: we read the full file anyway, so there is no waste there to use a large-ish chunk size, and it would kind of provide batching for prefetching "for free" (in the sense that if we want to prefetch, say, 128kb in advance, then we only need 2 chunks with a 64kb chunk, but need a lot more with a 4kb chunk size). --------- Co-authored-by: Branimir Lambov <branimir.lambov@datastax.com> (Rebase of commit c57974a)
michaelsembwever
pushed a commit
that referenced
this pull request
Feb 11, 2026
This introduces a `PrefetchingRebufferer` that, when enabled, prefetches (read from the underlying reader) a configurable number of chunks in advance (so only work on top of rebufferer factories that work with chunks, i.e. not uncompressed mmap). Prefetching must first be enabled globally by setting `-Dcassandra.read_prefetching_size_kb` to the desired value. With that, and assuming the disk access mode allows it (meaning, it's not uncompressed mmap), then prefetching will be applied to reads that are "clearly" sequential. Mostly, as of this patch, this means the sstable "scanners" and `SortedStringTableCursor`, so compaction, SAI index building and tools that read sstable fully (scrub, verifier) will benefit from it. Since rebufferers are synchronous, prefetching is done through a fixed thread pool. The number of thread of that pool can be set with `-Dcassandra.read_prefetching_threads` (but default to the number of "processors"). The `-Dcassandra.read_prefetching_window` can also be set to define how often prefetching is re-triggered. By default, when there is less than half of the value of `-Dcassandra.read_prefetching_size_kb` prefetched, then prefetching is triggered. See riptano/cndb#13639 for additional motivation. I'll note that this patch is adapted from the similar DSE behavior. However, there is a fair bit of modification since the DSE version was relying on the asynchronous nature of rebufferers there (it also had a few behavior I didn't fully understood and I didn't dig when it felt a bit specific to DSE). One point worth mentioning was that the DSE version was relying on the ability of the underlying channel to create batches when multiple chunks are prefetched. This is something that could have advantages and we could try to add back over time, but it's a lot harder to see how to do that in the non-async world of C*. Typically, in DSE, since all the prefetches were initiated on the current thread (without blocking it), building the "batch" was relatively easy, but it doesn't translate in this patch. Not to mention that currently the underlying channels have no batching options and those would have to be added (in a way that trickled down to the channel implementation within CNDB). Anyway, as of this patch, chunks are prefetched in parallel but with no batching optimization. Which does mean the `window' parameter is probably not as useful as for DSE, but I've kept it nonetheless as it's not exactly adding complexity. In the context of compaction in CNDB though, I think this mean that it would kind of be ideal if we could use a relatively large "chunk size" for the readers: we read the full file anyway, so there is no waste there to use a large-ish chunk size, and it would kind of provide batching for prefetching "for free" (in the sense that if we want to prefetch, say, 128kb in advance, then we only need 2 chunks with a 64kb chunk, but need a lot more with a 4kb chunk size). --------- Co-authored-by: Branimir Lambov <branimir.lambov@datastax.com> (Rebase of commit c57974a)
michaelsembwever
pushed a commit
that referenced
this pull request
Feb 12, 2026
This introduces a `PrefetchingRebufferer` that, when enabled, prefetches (read from the underlying reader) a configurable number of chunks in advance (so only work on top of rebufferer factories that work with chunks, i.e. not uncompressed mmap). Prefetching must first be enabled globally by setting `-Dcassandra.read_prefetching_size_kb` to the desired value. With that, and assuming the disk access mode allows it (meaning, it's not uncompressed mmap), then prefetching will be applied to reads that are "clearly" sequential. Mostly, as of this patch, this means the sstable "scanners" and `SortedStringTableCursor`, so compaction, SAI index building and tools that read sstable fully (scrub, verifier) will benefit from it. Since rebufferers are synchronous, prefetching is done through a fixed thread pool. The number of thread of that pool can be set with `-Dcassandra.read_prefetching_threads` (but default to the number of "processors"). The `-Dcassandra.read_prefetching_window` can also be set to define how often prefetching is re-triggered. By default, when there is less than half of the value of `-Dcassandra.read_prefetching_size_kb` prefetched, then prefetching is triggered. See riptano/cndb#13639 for additional motivation. I'll note that this patch is adapted from the similar DSE behavior. However, there is a fair bit of modification since the DSE version was relying on the asynchronous nature of rebufferers there (it also had a few behavior I didn't fully understood and I didn't dig when it felt a bit specific to DSE). One point worth mentioning was that the DSE version was relying on the ability of the underlying channel to create batches when multiple chunks are prefetched. This is something that could have advantages and we could try to add back over time, but it's a lot harder to see how to do that in the non-async world of C*. Typically, in DSE, since all the prefetches were initiated on the current thread (without blocking it), building the "batch" was relatively easy, but it doesn't translate in this patch. Not to mention that currently the underlying channels have no batching options and those would have to be added (in a way that trickled down to the channel implementation within CNDB). Anyway, as of this patch, chunks are prefetched in parallel but with no batching optimization. Which does mean the `window' parameter is probably not as useful as for DSE, but I've kept it nonetheless as it's not exactly adding complexity. In the context of compaction in CNDB though, I think this mean that it would kind of be ideal if we could use a relatively large "chunk size" for the readers: we read the full file anyway, so there is no waste there to use a large-ish chunk size, and it would kind of provide batching for prefetching "for free" (in the sense that if we want to prefetch, say, 128kb in advance, then we only need 2 chunks with a 64kb chunk, but need a lot more with a 4kb chunk size). --------- Co-authored-by: Branimir Lambov <branimir.lambov@datastax.com> (Rebase of commit c57974a)
michaelsembwever
pushed a commit
that referenced
this pull request
Feb 14, 2026
This introduces a `PrefetchingRebufferer` that, when enabled, prefetches (read from the underlying reader) a configurable number of chunks in advance (so only work on top of rebufferer factories that work with chunks, i.e. not uncompressed mmap). Prefetching must first be enabled globally by setting `-Dcassandra.read_prefetching_size_kb` to the desired value. With that, and assuming the disk access mode allows it (meaning, it's not uncompressed mmap), then prefetching will be applied to reads that are "clearly" sequential. Mostly, as of this patch, this means the sstable "scanners" and `SortedStringTableCursor`, so compaction, SAI index building and tools that read sstable fully (scrub, verifier) will benefit from it. Since rebufferers are synchronous, prefetching is done through a fixed thread pool. The number of thread of that pool can be set with `-Dcassandra.read_prefetching_threads` (but default to the number of "processors"). The `-Dcassandra.read_prefetching_window` can also be set to define how often prefetching is re-triggered. By default, when there is less than half of the value of `-Dcassandra.read_prefetching_size_kb` prefetched, then prefetching is triggered. See riptano/cndb#13639 for additional motivation. I'll note that this patch is adapted from the similar DSE behavior. However, there is a fair bit of modification since the DSE version was relying on the asynchronous nature of rebufferers there (it also had a few behavior I didn't fully understood and I didn't dig when it felt a bit specific to DSE). One point worth mentioning was that the DSE version was relying on the ability of the underlying channel to create batches when multiple chunks are prefetched. This is something that could have advantages and we could try to add back over time, but it's a lot harder to see how to do that in the non-async world of C*. Typically, in DSE, since all the prefetches were initiated on the current thread (without blocking it), building the "batch" was relatively easy, but it doesn't translate in this patch. Not to mention that currently the underlying channels have no batching options and those would have to be added (in a way that trickled down to the channel implementation within CNDB). Anyway, as of this patch, chunks are prefetched in parallel but with no batching optimization. Which does mean the `window' parameter is probably not as useful as for DSE, but I've kept it nonetheless as it's not exactly adding complexity. In the context of compaction in CNDB though, I think this mean that it would kind of be ideal if we could use a relatively large "chunk size" for the readers: we read the full file anyway, so there is no waste there to use a large-ish chunk size, and it would kind of provide batching for prefetching "for free" (in the sense that if we want to prefetch, say, 128kb in advance, then we only need 2 chunks with a 64kb chunk, but need a lot more with a 4kb chunk size). --------- Co-authored-by: Branimir Lambov <branimir.lambov@datastax.com> (Rebase of commit c57974a)
michaelsembwever
pushed a commit
that referenced
this pull request
Feb 16, 2026
This introduces a `PrefetchingRebufferer` that, when enabled, prefetches (read from the underlying reader) a configurable number of chunks in advance (so only work on top of rebufferer factories that work with chunks, i.e. not uncompressed mmap). Prefetching must first be enabled globally by setting `-Dcassandra.read_prefetching_size_kb` to the desired value. With that, and assuming the disk access mode allows it (meaning, it's not uncompressed mmap), then prefetching will be applied to reads that are "clearly" sequential. Mostly, as of this patch, this means the sstable "scanners" and `SortedStringTableCursor`, so compaction, SAI index building and tools that read sstable fully (scrub, verifier) will benefit from it. Since rebufferers are synchronous, prefetching is done through a fixed thread pool. The number of thread of that pool can be set with `-Dcassandra.read_prefetching_threads` (but default to the number of "processors"). The `-Dcassandra.read_prefetching_window` can also be set to define how often prefetching is re-triggered. By default, when there is less than half of the value of `-Dcassandra.read_prefetching_size_kb` prefetched, then prefetching is triggered. See riptano/cndb#13639 for additional motivation. I'll note that this patch is adapted from the similar DSE behavior. However, there is a fair bit of modification since the DSE version was relying on the asynchronous nature of rebufferers there (it also had a few behavior I didn't fully understood and I didn't dig when it felt a bit specific to DSE). One point worth mentioning was that the DSE version was relying on the ability of the underlying channel to create batches when multiple chunks are prefetched. This is something that could have advantages and we could try to add back over time, but it's a lot harder to see how to do that in the non-async world of C*. Typically, in DSE, since all the prefetches were initiated on the current thread (without blocking it), building the "batch" was relatively easy, but it doesn't translate in this patch. Not to mention that currently the underlying channels have no batching options and those would have to be added (in a way that trickled down to the channel implementation within CNDB). Anyway, as of this patch, chunks are prefetched in parallel but with no batching optimization. Which does mean the `window' parameter is probably not as useful as for DSE, but I've kept it nonetheless as it's not exactly adding complexity. In the context of compaction in CNDB though, I think this mean that it would kind of be ideal if we could use a relatively large "chunk size" for the readers: we read the full file anyway, so there is no waste there to use a large-ish chunk size, and it would kind of provide batching for prefetching "for free" (in the sense that if we want to prefetch, say, 128kb in advance, then we only need 2 chunks with a 64kb chunk, but need a lot more with a 4kb chunk size). --------- Co-authored-by: Branimir Lambov <branimir.lambov@datastax.com> (Rebase of commit c57974a)
michaelsembwever
pushed a commit
that referenced
this pull request
Feb 27, 2026
This introduces a `PrefetchingRebufferer` that, when enabled, prefetches (read from the underlying reader) a configurable number of chunks in advance (so only work on top of rebufferer factories that work with chunks, i.e. not uncompressed mmap). Prefetching must first be enabled globally by setting `-Dcassandra.read_prefetching_size_kb` to the desired value. With that, and assuming the disk access mode allows it (meaning, it's not uncompressed mmap), then prefetching will be applied to reads that are "clearly" sequential. Mostly, as of this patch, this means the sstable "scanners" and `SortedStringTableCursor`, so compaction, SAI index building and tools that read sstable fully (scrub, verifier) will benefit from it. Since rebufferers are synchronous, prefetching is done through a fixed thread pool. The number of thread of that pool can be set with `-Dcassandra.read_prefetching_threads` (but default to the number of "processors"). The `-Dcassandra.read_prefetching_window` can also be set to define how often prefetching is re-triggered. By default, when there is less than half of the value of `-Dcassandra.read_prefetching_size_kb` prefetched, then prefetching is triggered. See riptano/cndb#13639 for additional motivation. I'll note that this patch is adapted from the similar DSE behavior. However, there is a fair bit of modification since the DSE version was relying on the asynchronous nature of rebufferers there (it also had a few behavior I didn't fully understood and I didn't dig when it felt a bit specific to DSE). One point worth mentioning was that the DSE version was relying on the ability of the underlying channel to create batches when multiple chunks are prefetched. This is something that could have advantages and we could try to add back over time, but it's a lot harder to see how to do that in the non-async world of C*. Typically, in DSE, since all the prefetches were initiated on the current thread (without blocking it), building the "batch" was relatively easy, but it doesn't translate in this patch. Not to mention that currently the underlying channels have no batching options and those would have to be added (in a way that trickled down to the channel implementation within CNDB). Anyway, as of this patch, chunks are prefetched in parallel but with no batching optimization. Which does mean the `window' parameter is probably not as useful as for DSE, but I've kept it nonetheless as it's not exactly adding complexity. In the context of compaction in CNDB though, I think this mean that it would kind of be ideal if we could use a relatively large "chunk size" for the readers: we read the full file anyway, so there is no waste there to use a large-ish chunk size, and it would kind of provide batching for prefetching "for free" (in the sense that if we want to prefetch, say, 128kb in advance, then we only need 2 chunks with a 64kb chunk, but need a lot more with a 4kb chunk size). --------- Co-authored-by: Branimir Lambov <branimir.lambov@datastax.com> (Rebase of commit c57974a)
michaelsembwever
pushed a commit
that referenced
this pull request
Mar 2, 2026
This introduces a `PrefetchingRebufferer` that, when enabled, prefetches (read from the underlying reader) a configurable number of chunks in advance (so only work on top of rebufferer factories that work with chunks, i.e. not uncompressed mmap). Prefetching must first be enabled globally by setting `-Dcassandra.read_prefetching_size_kb` to the desired value. With that, and assuming the disk access mode allows it (meaning, it's not uncompressed mmap), then prefetching will be applied to reads that are "clearly" sequential. Mostly, as of this patch, this means the sstable "scanners" and `SortedStringTableCursor`, so compaction, SAI index building and tools that read sstable fully (scrub, verifier) will benefit from it. Since rebufferers are synchronous, prefetching is done through a fixed thread pool. The number of thread of that pool can be set with `-Dcassandra.read_prefetching_threads` (but default to the number of "processors"). The `-Dcassandra.read_prefetching_window` can also be set to define how often prefetching is re-triggered. By default, when there is less than half of the value of `-Dcassandra.read_prefetching_size_kb` prefetched, then prefetching is triggered. See riptano/cndb#13639 for additional motivation. I'll note that this patch is adapted from the similar DSE behavior. However, there is a fair bit of modification since the DSE version was relying on the asynchronous nature of rebufferers there (it also had a few behavior I didn't fully understood and I didn't dig when it felt a bit specific to DSE). One point worth mentioning was that the DSE version was relying on the ability of the underlying channel to create batches when multiple chunks are prefetched. This is something that could have advantages and we could try to add back over time, but it's a lot harder to see how to do that in the non-async world of C*. Typically, in DSE, since all the prefetches were initiated on the current thread (without blocking it), building the "batch" was relatively easy, but it doesn't translate in this patch. Not to mention that currently the underlying channels have no batching options and those would have to be added (in a way that trickled down to the channel implementation within CNDB). Anyway, as of this patch, chunks are prefetched in parallel but with no batching optimization. Which does mean the `window' parameter is probably not as useful as for DSE, but I've kept it nonetheless as it's not exactly adding complexity. In the context of compaction in CNDB though, I think this mean that it would kind of be ideal if we could use a relatively large "chunk size" for the readers: we read the full file anyway, so there is no waste there to use a large-ish chunk size, and it would kind of provide batching for prefetching "for free" (in the sense that if we want to prefetch, say, 128kb in advance, then we only need 2 chunks with a 64kb chunk, but need a lot more with a 4kb chunk size). --------- Co-authored-by: Branimir Lambov <branimir.lambov@datastax.com> (Rebase of commit c57974a)
michaelsembwever
pushed a commit
that referenced
this pull request
Mar 4, 2026
This introduces a `PrefetchingRebufferer` that, when enabled, prefetches (read from the underlying reader) a configurable number of chunks in advance (so only work on top of rebufferer factories that work with chunks, i.e. not uncompressed mmap). Prefetching must first be enabled globally by setting `-Dcassandra.read_prefetching_size_kb` to the desired value. With that, and assuming the disk access mode allows it (meaning, it's not uncompressed mmap), then prefetching will be applied to reads that are "clearly" sequential. Mostly, as of this patch, this means the sstable "scanners" and `SortedStringTableCursor`, so compaction, SAI index building and tools that read sstable fully (scrub, verifier) will benefit from it. Since rebufferers are synchronous, prefetching is done through a fixed thread pool. The number of thread of that pool can be set with `-Dcassandra.read_prefetching_threads` (but default to the number of "processors"). The `-Dcassandra.read_prefetching_window` can also be set to define how often prefetching is re-triggered. By default, when there is less than half of the value of `-Dcassandra.read_prefetching_size_kb` prefetched, then prefetching is triggered. See riptano/cndb#13639 for additional motivation. I'll note that this patch is adapted from the similar DSE behavior. However, there is a fair bit of modification since the DSE version was relying on the asynchronous nature of rebufferers there (it also had a few behavior I didn't fully understood and I didn't dig when it felt a bit specific to DSE). One point worth mentioning was that the DSE version was relying on the ability of the underlying channel to create batches when multiple chunks are prefetched. This is something that could have advantages and we could try to add back over time, but it's a lot harder to see how to do that in the non-async world of C*. Typically, in DSE, since all the prefetches were initiated on the current thread (without blocking it), building the "batch" was relatively easy, but it doesn't translate in this patch. Not to mention that currently the underlying channels have no batching options and those would have to be added (in a way that trickled down to the channel implementation within CNDB). Anyway, as of this patch, chunks are prefetched in parallel but with no batching optimization. Which does mean the `window' parameter is probably not as useful as for DSE, but I've kept it nonetheless as it's not exactly adding complexity. In the context of compaction in CNDB though, I think this mean that it would kind of be ideal if we could use a relatively large "chunk size" for the readers: we read the full file anyway, so there is no waste there to use a large-ish chunk size, and it would kind of provide batching for prefetching "for free" (in the sense that if we want to prefetch, say, 128kb in advance, then we only need 2 chunks with a 64kb chunk, but need a lot more with a 4kb chunk size). --------- Co-authored-by: Branimir Lambov <branimir.lambov@datastax.com> (Rebase of commit c57974a)
michaelsembwever
pushed a commit
that referenced
this pull request
Mar 5, 2026
This introduces a `PrefetchingRebufferer` that, when enabled, prefetches (read from the underlying reader) a configurable number of chunks in advance (so only work on top of rebufferer factories that work with chunks, i.e. not uncompressed mmap). Prefetching must first be enabled globally by setting `-Dcassandra.read_prefetching_size_kb` to the desired value. With that, and assuming the disk access mode allows it (meaning, it's not uncompressed mmap), then prefetching will be applied to reads that are "clearly" sequential. Mostly, as of this patch, this means the sstable "scanners" and `SortedStringTableCursor`, so compaction, SAI index building and tools that read sstable fully (scrub, verifier) will benefit from it. Since rebufferers are synchronous, prefetching is done through a fixed thread pool. The number of thread of that pool can be set with `-Dcassandra.read_prefetching_threads` (but default to the number of "processors"). The `-Dcassandra.read_prefetching_window` can also be set to define how often prefetching is re-triggered. By default, when there is less than half of the value of `-Dcassandra.read_prefetching_size_kb` prefetched, then prefetching is triggered. See riptano/cndb#13639 for additional motivation. I'll note that this patch is adapted from the similar DSE behavior. However, there is a fair bit of modification since the DSE version was relying on the asynchronous nature of rebufferers there (it also had a few behavior I didn't fully understood and I didn't dig when it felt a bit specific to DSE). One point worth mentioning was that the DSE version was relying on the ability of the underlying channel to create batches when multiple chunks are prefetched. This is something that could have advantages and we could try to add back over time, but it's a lot harder to see how to do that in the non-async world of C*. Typically, in DSE, since all the prefetches were initiated on the current thread (without blocking it), building the "batch" was relatively easy, but it doesn't translate in this patch. Not to mention that currently the underlying channels have no batching options and those would have to be added (in a way that trickled down to the channel implementation within CNDB). Anyway, as of this patch, chunks are prefetched in parallel but with no batching optimization. Which does mean the `window' parameter is probably not as useful as for DSE, but I've kept it nonetheless as it's not exactly adding complexity. In the context of compaction in CNDB though, I think this mean that it would kind of be ideal if we could use a relatively large "chunk size" for the readers: we read the full file anyway, so there is no waste there to use a large-ish chunk size, and it would kind of provide batching for prefetching "for free" (in the sense that if we want to prefetch, say, 128kb in advance, then we only need 2 chunks with a 64kb chunk, but need a lot more with a 4kb chunk size). --------- Co-authored-by: Branimir Lambov <branimir.lambov@datastax.com> (Rebase of commit c57974a)
michaelsembwever
pushed a commit
that referenced
this pull request
Mar 25, 2026
This introduces a `PrefetchingRebufferer` that, when enabled, prefetches (read from the underlying reader) a configurable number of chunks in advance (so only work on top of rebufferer factories that work with chunks, i.e. not uncompressed mmap). Prefetching must first be enabled globally by setting `-Dcassandra.read_prefetching_size_kb` to the desired value. With that, and assuming the disk access mode allows it (meaning, it's not uncompressed mmap), then prefetching will be applied to reads that are "clearly" sequential. Mostly, as of this patch, this means the sstable "scanners" and `SortedStringTableCursor`, so compaction, SAI index building and tools that read sstable fully (scrub, verifier) will benefit from it. Since rebufferers are synchronous, prefetching is done through a fixed thread pool. The number of thread of that pool can be set with `-Dcassandra.read_prefetching_threads` (but default to the number of "processors"). The `-Dcassandra.read_prefetching_window` can also be set to define how often prefetching is re-triggered. By default, when there is less than half of the value of `-Dcassandra.read_prefetching_size_kb` prefetched, then prefetching is triggered. See riptano/cndb#13639 for additional motivation. I'll note that this patch is adapted from the similar DSE behavior. However, there is a fair bit of modification since the DSE version was relying on the asynchronous nature of rebufferers there (it also had a few behavior I didn't fully understood and I didn't dig when it felt a bit specific to DSE). One point worth mentioning was that the DSE version was relying on the ability of the underlying channel to create batches when multiple chunks are prefetched. This is something that could have advantages and we could try to add back over time, but it's a lot harder to see how to do that in the non-async world of C*. Typically, in DSE, since all the prefetches were initiated on the current thread (without blocking it), building the "batch" was relatively easy, but it doesn't translate in this patch. Not to mention that currently the underlying channels have no batching options and those would have to be added (in a way that trickled down to the channel implementation within CNDB). Anyway, as of this patch, chunks are prefetched in parallel but with no batching optimization. Which does mean the `window' parameter is probably not as useful as for DSE, but I've kept it nonetheless as it's not exactly adding complexity. In the context of compaction in CNDB though, I think this mean that it would kind of be ideal if we could use a relatively large "chunk size" for the readers: we read the full file anyway, so there is no waste there to use a large-ish chunk size, and it would kind of provide batching for prefetching "for free" (in the sense that if we want to prefetch, say, 128kb in advance, then we only need 2 chunks with a 64kb chunk, but need a lot more with a 4kb chunk size). --------- Co-authored-by: Branimir Lambov <branimir.lambov@datastax.com> (Rebase of commit c57974a)
michaelsembwever
pushed a commit
that referenced
this pull request
Mar 27, 2026
This introduces a `PrefetchingRebufferer` that, when enabled, prefetches (read from the underlying reader) a configurable number of chunks in advance (so only work on top of rebufferer factories that work with chunks, i.e. not uncompressed mmap). Prefetching must first be enabled globally by setting `-Dcassandra.read_prefetching_size_kb` to the desired value. With that, and assuming the disk access mode allows it (meaning, it's not uncompressed mmap), then prefetching will be applied to reads that are "clearly" sequential. Mostly, as of this patch, this means the sstable "scanners" and `SortedStringTableCursor`, so compaction, SAI index building and tools that read sstable fully (scrub, verifier) will benefit from it. Since rebufferers are synchronous, prefetching is done through a fixed thread pool. The number of thread of that pool can be set with `-Dcassandra.read_prefetching_threads` (but default to the number of "processors"). The `-Dcassandra.read_prefetching_window` can also be set to define how often prefetching is re-triggered. By default, when there is less than half of the value of `-Dcassandra.read_prefetching_size_kb` prefetched, then prefetching is triggered. See riptano/cndb#13639 for additional motivation. I'll note that this patch is adapted from the similar DSE behavior. However, there is a fair bit of modification since the DSE version was relying on the asynchronous nature of rebufferers there (it also had a few behavior I didn't fully understood and I didn't dig when it felt a bit specific to DSE). One point worth mentioning was that the DSE version was relying on the ability of the underlying channel to create batches when multiple chunks are prefetched. This is something that could have advantages and we could try to add back over time, but it's a lot harder to see how to do that in the non-async world of C*. Typically, in DSE, since all the prefetches were initiated on the current thread (without blocking it), building the "batch" was relatively easy, but it doesn't translate in this patch. Not to mention that currently the underlying channels have no batching options and those would have to be added (in a way that trickled down to the channel implementation within CNDB). Anyway, as of this patch, chunks are prefetched in parallel but with no batching optimization. Which does mean the `window' parameter is probably not as useful as for DSE, but I've kept it nonetheless as it's not exactly adding complexity. In the context of compaction in CNDB though, I think this mean that it would kind of be ideal if we could use a relatively large "chunk size" for the readers: we read the full file anyway, so there is no waste there to use a large-ish chunk size, and it would kind of provide batching for prefetching "for free" (in the sense that if we want to prefetch, say, 128kb in advance, then we only need 2 chunks with a 64kb chunk, but need a lot more with a 4kb chunk size). --------- Co-authored-by: Branimir Lambov <branimir.lambov@datastax.com> (Rebase of commit c57974a)
michaelsembwever
pushed a commit
that referenced
this pull request
Apr 14, 2026
This introduces a `PrefetchingRebufferer` that, when enabled, prefetches (read from the underlying reader) a configurable number of chunks in advance (so only work on top of rebufferer factories that work with chunks, i.e. not uncompressed mmap). Prefetching must first be enabled globally by setting `-Dcassandra.read_prefetching_size_kb` to the desired value. With that, and assuming the disk access mode allows it (meaning, it's not uncompressed mmap), then prefetching will be applied to reads that are "clearly" sequential. Mostly, as of this patch, this means the sstable "scanners" and `SortedStringTableCursor`, so compaction, SAI index building and tools that read sstable fully (scrub, verifier) will benefit from it. Since rebufferers are synchronous, prefetching is done through a fixed thread pool. The number of thread of that pool can be set with `-Dcassandra.read_prefetching_threads` (but default to the number of "processors"). The `-Dcassandra.read_prefetching_window` can also be set to define how often prefetching is re-triggered. By default, when there is less than half of the value of `-Dcassandra.read_prefetching_size_kb` prefetched, then prefetching is triggered. See riptano/cndb#13639 for additional motivation. I'll note that this patch is adapted from the similar DSE behavior. However, there is a fair bit of modification since the DSE version was relying on the asynchronous nature of rebufferers there (it also had a few behavior I didn't fully understood and I didn't dig when it felt a bit specific to DSE). One point worth mentioning was that the DSE version was relying on the ability of the underlying channel to create batches when multiple chunks are prefetched. This is something that could have advantages and we could try to add back over time, but it's a lot harder to see how to do that in the non-async world of C*. Typically, in DSE, since all the prefetches were initiated on the current thread (without blocking it), building the "batch" was relatively easy, but it doesn't translate in this patch. Not to mention that currently the underlying channels have no batching options and those would have to be added (in a way that trickled down to the channel implementation within CNDB). Anyway, as of this patch, chunks are prefetched in parallel but with no batching optimization. Which does mean the `window' parameter is probably not as useful as for DSE, but I've kept it nonetheless as it's not exactly adding complexity. In the context of compaction in CNDB though, I think this mean that it would kind of be ideal if we could use a relatively large "chunk size" for the readers: we read the full file anyway, so there is no waste there to use a large-ish chunk size, and it would kind of provide batching for prefetching "for free" (in the sense that if we want to prefetch, say, 128kb in advance, then we only need 2 chunks with a 64kb chunk, but need a lot more with a 4kb chunk size). --------- Co-authored-by: Branimir Lambov <branimir.lambov@datastax.com> (Rebase of commit c57974a)
michaelsembwever
pushed a commit
that referenced
this pull request
Apr 15, 2026
This introduces a `PrefetchingRebufferer` that, when enabled, prefetches (read from the underlying reader) a configurable number of chunks in advance (so only work on top of rebufferer factories that work with chunks, i.e. not uncompressed mmap). Prefetching must first be enabled globally by setting `-Dcassandra.read_prefetching_size_kb` to the desired value. With that, and assuming the disk access mode allows it (meaning, it's not uncompressed mmap), then prefetching will be applied to reads that are "clearly" sequential. Mostly, as of this patch, this means the sstable "scanners" and `SortedStringTableCursor`, so compaction, SAI index building and tools that read sstable fully (scrub, verifier) will benefit from it. Since rebufferers are synchronous, prefetching is done through a fixed thread pool. The number of thread of that pool can be set with `-Dcassandra.read_prefetching_threads` (but default to the number of "processors"). The `-Dcassandra.read_prefetching_window` can also be set to define how often prefetching is re-triggered. By default, when there is less than half of the value of `-Dcassandra.read_prefetching_size_kb` prefetched, then prefetching is triggered. See riptano/cndb#13639 for additional motivation. I'll note that this patch is adapted from the similar DSE behavior. However, there is a fair bit of modification since the DSE version was relying on the asynchronous nature of rebufferers there (it also had a few behavior I didn't fully understood and I didn't dig when it felt a bit specific to DSE). One point worth mentioning was that the DSE version was relying on the ability of the underlying channel to create batches when multiple chunks are prefetched. This is something that could have advantages and we could try to add back over time, but it's a lot harder to see how to do that in the non-async world of C*. Typically, in DSE, since all the prefetches were initiated on the current thread (without blocking it), building the "batch" was relatively easy, but it doesn't translate in this patch. Not to mention that currently the underlying channels have no batching options and those would have to be added (in a way that trickled down to the channel implementation within CNDB). Anyway, as of this patch, chunks are prefetched in parallel but with no batching optimization. Which does mean the `window' parameter is probably not as useful as for DSE, but I've kept it nonetheless as it's not exactly adding complexity. In the context of compaction in CNDB though, I think this mean that it would kind of be ideal if we could use a relatively large "chunk size" for the readers: we read the full file anyway, so there is no waste there to use a large-ish chunk size, and it would kind of provide batching for prefetching "for free" (in the sense that if we want to prefetch, say, 128kb in advance, then we only need 2 chunks with a 64kb chunk, but need a lot more with a 4kb chunk size). --------- Co-authored-by: Branimir Lambov <branimir.lambov@datastax.com> (Rebase of commit c57974a)
This file contains hidden or 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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
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.
This introduces a
PrefetchingRebuffererthat, when enabled, prefetches (read from the underlying reader) a configurable number of chunks in advance (so only work on top of rebufferer factories that work with chunks, i.e. not uncompressed mmap).Prefetching must first be enabled globally by setting
-Dcassandra.read_prefetching_size_kbto the desired value. With that, and assuming the disk access mode allows it (meaning, it's not uncompressed mmap), then prefetching will be applied to reads that are "clearly" sequential. Mostly, as of this patch, this means the sstable "scanners" andSortedStringTableCursor, so compaction, SAI index building and tools that read sstable fully (scrub, verifier) will benefit from it.Since rebufferers are synchronous, prefetching is done through a fixed thread pool. The number of thread of that pool can be set with
-Dcassandra.read_prefetching_threads(but default to the number of "processors").The
-Dcassandra.read_prefetching_windowcan also be set to define how often prefetching is re-triggered. By default, when there is less than half of the value of-Dcassandra.read_prefetching_size_kbprefetched, then prefetching is triggered.See https://github.com/riptano/cndb/issues/13639 for additional motivation.
I'll note that this patch is adapted from the similar DSE behavior. However, there is a fair bit of modification since the DSE version was relying on the asynchronous nature of rebufferers there (it also had a few behavior I didn't fully understood and I didn't dig when it felt a bit specific to DSE).
One point worth mentioning was that the DSE version was relying on the ability of the underlying channel to create batches when multiple chunks are prefetched. This is something that could have advantages and we could try to add back over time, but it's a lot harder to see how to do that in the non-async world of C*. Typically, in DSE, since all the prefetches were initiated on the current thread (without blocking it), building the "batch" was relatively easy, but it doesn't translate in this patch. Not to mention that currently the underlying channels have no batching options and those would have to be added (in a way that trickled down to the channel implementation within CNDB). Anyway, as of this patch, chunks are prefetched in parallel but with no batching optimization. Which does mean the `window' parameter is probably not as useful as for DSE, but I've kept it nonetheless as it's not exactly adding complexity.
In the context of compaction in CNDB though, I think this mean that it would kind of be ideal if we could use a relatively large "chunk size" for the readers: we read the full file anyway, so there is no waste there to use a large-ish chunk size, and it would kind of provide batching for prefetching "for free" (in the sense that if we want to prefetch, say, 128kb in advance, then we only need 2 chunks with a 64kb chunk, but need a lot more with a 4kb chunk size).