kv: re-enable time-bound iterators for RefreshRange request

Closes cockroachdb#53348. The other two requests mentioned in that issue (`ResolveIntentRange` and `EndTxn`) would no longer benefit from time-bound iterators because, thanks to b5213fd, they no longer scan the MVCC keyspace. Transaction refreshing is a form of an optimistic concurrency control validation phase. Before a transaction can commit, if it will be committing at a timestamp higher than its original timestamp, it issues point and ranged refresh requests to the key spans it had previous read. The refresh requests scan a span of keys and determine whether any new values have been written since the transaction originally read the keys. The use of time-bound iterators is an important optimization for ranged refresh operations because we expect very few new writes between the time that a transaction originally reads and the time that it refreshes. Without this optimization, each refresh was redoing all of a transaction's reads at their original cost. This effectively doubled the cost of reads for transactions that had to refresh (or worse for those that refreshed multiple times). With this optimization, refreshing a span of keys is expected to be significantly cheaper than the original scan over that span of keys, because it can ignore most files in the lower levels of the LSM. RefreshRange requests were originally built to use time-bound iterators. However, this optimization was disabled in 1eb3b2a due to concerns about correctness. Since then, then correctness concerns have been addressed and we have begun using time-bound iterators in a handful of places. This commit re-enables time-bound iterators for `RefreshRange` requests. It does so by using `MVCCIncrementalIterator`, which was enhanced to support additional "intent policies" in 87c7f11. This commit uses the "emit" intent policy so that `RefreshRange` will observe all values and all intents in the refresh time window. ---- Microbenchmarks: ``` name old time/op new time/op delta RefreshRange/linear-keys/refresh_window=[95.00,99.00]-10 230ms ± 1% 0ms ± 1% -99.99% (p=0.000 n=9+9) RefreshRange/linear-keys/refresh_window=[75.00,99.00]-10 185ms ± 1% 0ms ± 1% -99.99% (p=0.000 n=10+10) RefreshRange/linear-keys/refresh_window=[75.00,95.00]-10 185ms ± 1% 0ms ± 1% -99.99% (p=0.000 n=9+10) RefreshRange/linear-keys/refresh_window=[50.00,75.00]-10 123ms ± 1% 0ms ± 2% -99.99% (p=0.000 n=10+10) RefreshRange/linear-keys/refresh_window=[50.00,95.00]-10 123ms ± 1% 0ms ± 2% -99.99% (p=0.000 n=10+10) RefreshRange/linear-keys/refresh_window=[50.00,99.00]-10 123ms ± 0% 0ms ± 1% -99.99% (p=0.000 n=10+8) RefreshRange/linear-keys/refresh_window=[99.00,99.00]-10 240ms ± 1% 0ms ± 1% -99.96% (p=0.000 n=9+10) RefreshRange/linear-keys/refresh_window=[95.00,95.00]-10 237ms ± 1% 0ms ± 2% -99.96% (p=0.000 n=10+10) RefreshRange/linear-keys/refresh_window=[75.00,75.00]-10 224ms ± 0% 0ms ± 1% -99.95% (p=0.000 n=9+9) RefreshRange/linear-keys/refresh_window=[50.00,50.00]-10 207ms ± 1% 0ms ± 3% -99.95% (p=0.000 n=10+10) RefreshRange/linear-keys/refresh_window=[0.00,0.00]-10 174ms ± 1% 0ms ± 1% -99.93% (p=0.000 n=10+10) RefreshRange/random-keys/refresh_window=[0.00,0.00]-10 189ms ± 1% 0ms ± 1% -99.86% (p=0.000 n=9+9) RefreshRange/mixed-case/refresh_window=[0.00,0.00]-10 184ms ± 0% 0ms ± 0% -99.85% (p=0.000 n=8+9) RefreshRange/mixed-case/refresh_window=[95.00,95.00]-10 252ms ± 0% 1ms ± 2% -99.70% (p=0.000 n=8+10) RefreshRange/random-keys/refresh_window=[0.00,50.00]-10 412µs ± 1% 13µs ± 2% -96.83% (p=0.000 n=10+10) RefreshRange/mixed-case/refresh_window=[0.00,50.00]-10 413µs ± 1% 13µs ± 1% -96.78% (p=0.000 n=10+10) RefreshRange/mixed-case/refresh_window=[0.00,75.00]-10 292µs ± 1% 13µs ± 1% -95.43% (p=0.000 n=10+10) RefreshRange/random-keys/refresh_window=[0.00,95.00]-10 245µs ± 1% 13µs ± 2% -94.64% (p=0.000 n=10+10) RefreshRange/random-keys/refresh_window=[50.00,95.00]-10 245µs ± 0% 14µs ± 1% -94.49% (p=0.000 n=9+10) RefreshRange/random-keys/refresh_window=[0.00,99.00]-10 238µs ± 1% 13µs ± 1% -94.48% (p=0.000 n=10+10) RefreshRange/mixed-case/refresh_window=[95.00,99.00]-10 237µs ± 1% 13µs ± 2% -94.42% (p=0.000 n=10+10) RefreshRange/random-keys/refresh_window=[50.00,99.00]-10 237µs ± 2% 14µs ± 1% -94.29% (p=0.000 n=9+10) RefreshRange/mixed-case/refresh_window=[50.00,75.00]-10 292µs ± 1% 17µs ± 1% -94.07% (p=0.000 n=10+9) RefreshRange/linear-keys/refresh_window=[0.00,75.00]-10 225µs ± 2% 14µs ± 1% -94.00% (p=0.000 n=10+10) RefreshRange/linear-keys/refresh_window=[0.00,99.00]-10 224µs ± 1% 13µs ± 1% -93.99% (p=0.000 n=10+8) RefreshRange/linear-keys/refresh_window=[0.00,95.00]-10 224µs ± 1% 14µs ± 1% -93.95% (p=0.000 n=10+10) RefreshRange/random-keys/refresh_window=[75.00,95.00]-10 244µs ± 0% 15µs ± 1% -93.86% (p=0.000 n=7+10) RefreshRange/mixed-case/refresh_window=[0.00,99.00]-10 237µs ± 1% 15µs ± 1% -93.82% (p=0.000 n=10+10) RefreshRange/linear-keys/refresh_window=[0.00,50.00]-10 224µs ± 1% 14µs ± 1% -93.76% (p=0.000 n=10+10) RefreshRange/mixed-case/refresh_window=[0.00,95.00]-10 244µs ± 1% 15µs ± 1% -93.75% (p=0.000 n=10+10) RefreshRange/random-keys/refresh_window=[75.00,99.00]-10 238µs ± 1% 15µs ± 1% -93.72% (p=0.000 n=9+10) RefreshRange/mixed-case/refresh_window=[75.00,99.00]-10 236µs ± 1% 15µs ± 2% -93.64% (p=0.000 n=10+10) RefreshRange/mixed-case/refresh_window=[50.00,99.00]-10 236µs ± 1% 15µs ± 1% -93.63% (p=0.000 n=10+9) RefreshRange/mixed-case/refresh_window=[50.00,95.00]-10 244µs ± 0% 16µs ± 1% -93.58% (p=0.000 n=10+9) RefreshRange/mixed-case/refresh_window=[75.00,95.00]-10 244µs ± 1% 16µs ± 0% -93.55% (p=0.000 n=10+8) RefreshRange/random-keys/refresh_window=[0.00,75.00]-10 287µs ± 1% 19µs ± 1% -93.20% (p=0.000 n=10+10) RefreshRange/random-keys/refresh_window=[95.00,99.00]-10 237µs ± 1% 17µs ± 1% -92.69% (p=0.000 n=10+10) RefreshRange/random-keys/refresh_window=[50.00,75.00]-10 288µs ± 2% 23µs ± 1% -91.98% (p=0.000 n=9+10) RefreshRange/mixed-case/refresh_window=[99.00,99.00]-10 255ms ± 1% 122ms ± 1% -52.02% (p=0.000 n=9+9) RefreshRange/random-keys/refresh_window=[75.00,75.00]-10 242ms ± 1% 152ms ± 1% -37.02% (p=0.000 n=10+9) RefreshRange/random-keys/refresh_window=[99.00,99.00]-10 259ms ± 0% 354ms ± 1% +36.73% (p=0.000 n=7+9) RefreshRange/random-keys/refresh_window=[95.00,95.00]-10 256ms ± 1% 353ms ± 1% +37.65% (p=0.000 n=10+9) RefreshRange/mixed-case/refresh_window=[75.00,75.00]-10 242ms ± 0% 398ms ± 1% +64.38% (p=0.000 n=9+10) RefreshRange/mixed-case/refresh_window=[50.00,50.00]-10 227ms ± 0% 392ms ± 1% +72.65% (p=0.000 n=9+10) RefreshRange/random-keys/refresh_window=[50.00,50.00]-10 229ms ± 1% 512ms ± 1% +123.45% (p=0.000 n=9+9) name old alloc/op new alloc/op delta RefreshRange/linear-keys/refresh_window=[99.00,99.00]-10 195MB ± 0% 0MB ± 0% -100.00% (p=0.000 n=9+10) RefreshRange/linear-keys/refresh_window=[95.00,95.00]-10 188MB ± 0% 0MB ± 0% -100.00% (p=0.000 n=10+10) RefreshRange/mixed-case/refresh_window=[95.00,95.00]-10 188MB ± 0% 0MB ± 1% -100.00% (p=0.000 n=10+10) RefreshRange/linear-keys/refresh_window=[75.00,75.00]-10 148MB ± 0% 0MB ± 0% -100.00% (p=0.000 n=10+10) RefreshRange/linear-keys/refresh_window=[50.00,50.00]-10 98.8MB ± 0% 0.0MB ± 0% -100.00% (p=0.000 n=10+10) RefreshRange/mixed-case/refresh_window=[99.00,99.00]-10 195MB ± 0% 0MB ± 3% -100.00% (p=0.000 n=9+8) RefreshRange/linear-keys/refresh_window=[95.00,99.00]-10 188MB ± 0% 0MB ± 0% -100.00% (p=0.000 n=10+10) RefreshRange/linear-keys/refresh_window=[75.00,95.00]-10 148MB ± 0% 0MB ± 0% -99.99% (p=0.000 n=10+10) RefreshRange/linear-keys/refresh_window=[75.00,99.00]-10 148MB ± 0% 0MB ± 0% -99.99% (p=0.000 n=9+10) RefreshRange/linear-keys/refresh_window=[50.00,75.00]-10 99.0MB ± 0% 0.0MB ± 0% -99.99% (p=0.000 n=9+10) RefreshRange/linear-keys/refresh_window=[50.00,95.00]-10 99.0MB ± 0% 0.0MB ± 0% -99.99% (p=0.000 n=8+9) RefreshRange/linear-keys/refresh_window=[50.00,99.00]-10 99.0MB ± 0% 0.0MB ± 0% -99.99% (p=0.000 n=9+10) RefreshRange/mixed-case/refresh_window=[75.00,75.00]-10 148MB ± 0% 0MB ±29% -99.99% (p=0.000 n=10+10) RefreshRange/random-keys/refresh_window=[75.00,75.00]-10 148MB ± 0% 0MB ± 6% -99.98% (p=0.000 n=10+9) RefreshRange/mixed-case/refresh_window=[50.00,50.00]-10 98.7MB ± 0% 0.0MB ± 1% -99.98% (p=0.000 n=10+8) RefreshRange/random-keys/refresh_window=[99.00,99.00]-10 195MB ± 0% 0MB ± 4% -99.97% (p=0.000 n=9+8) RefreshRange/random-keys/refresh_window=[95.00,95.00]-10 188MB ± 0% 0MB ± 3% -99.97% (p=0.000 n=10+8) RefreshRange/random-keys/refresh_window=[50.00,50.00]-10 98.7MB ± 0% 0.1MB ±12% -99.92% (p=0.000 n=10+10) RefreshRange/linear-keys/refresh_window=[0.00,0.00]-10 41.9kB ± 5% 1.1kB ± 0% -97.27% (p=0.000 n=9+9) RefreshRange/linear-keys/refresh_window=[0.00,50.00]-10 208kB ± 0% 8kB ± 0% -96.04% (p=0.000 n=10+10) RefreshRange/mixed-case/refresh_window=[95.00,99.00]-10 208kB ± 0% 8kB ± 0% -96.03% (p=0.000 n=10+10) RefreshRange/random-keys/refresh_window=[95.00,99.00]-10 208kB ± 0% 8kB ± 0% -96.03% (p=0.000 n=10+9) RefreshRange/mixed-case/refresh_window=[50.00,75.00]-10 208kB ± 0% 8kB ± 0% -96.03% (p=0.000 n=8+10) RefreshRange/mixed-case/refresh_window=[0.00,75.00]-10 208kB ± 0% 8kB ± 0% -96.03% (p=0.000 n=10+9) RefreshRange/mixed-case/refresh_window=[0.00,50.00]-10 207kB ± 0% 8kB ± 0% -96.03% (p=0.000 n=9+10) RefreshRange/linear-keys/refresh_window=[0.00,95.00]-10 208kB ± 0% 8kB ± 0% -96.03% (p=0.000 n=10+10) RefreshRange/random-keys/refresh_window=[0.00,50.00]-10 207kB ± 0% 8kB ± 0% -96.03% (p=0.000 n=10+10) RefreshRange/linear-keys/refresh_window=[0.00,75.00]-10 208kB ± 0% 8kB ± 0% -96.02% (p=0.000 n=10+10) RefreshRange/linear-keys/refresh_window=[0.00,99.00]-10 208kB ± 0% 8kB ± 0% -96.02% (p=0.000 n=10+10) RefreshRange/random-keys/refresh_window=[0.00,99.00]-10 208kB ± 0% 8kB ± 0% -96.02% (p=0.000 n=10+10) RefreshRange/random-keys/refresh_window=[50.00,99.00]-10 208kB ± 0% 8kB ± 0% -96.02% (p=0.000 n=10+10) RefreshRange/random-keys/refresh_window=[50.00,95.00]-10 208kB ± 0% 8kB ± 0% -96.02% (p=0.000 n=9+10) RefreshRange/random-keys/refresh_window=[75.00,99.00]-10 208kB ± 0% 8kB ± 0% -96.02% (p=0.000 n=10+10) RefreshRange/random-keys/refresh_window=[0.00,95.00]-10 208kB ± 0% 8kB ± 0% -96.02% (p=0.000 n=10+10) RefreshRange/random-keys/refresh_window=[75.00,95.00]-10 208kB ± 0% 8kB ± 0% -96.02% (p=0.000 n=10+10) RefreshRange/random-keys/refresh_window=[0.00,75.00]-10 207kB ± 0% 8kB ± 0% -96.02% (p=0.000 n=10+10) RefreshRange/random-keys/refresh_window=[50.00,75.00]-10 207kB ± 0% 8kB ± 0% -96.02% (p=0.000 n=9+10) RefreshRange/mixed-case/refresh_window=[0.00,99.00]-10 208kB ± 0% 8kB ± 0% -96.01% (p=0.000 n=10+10) RefreshRange/mixed-case/refresh_window=[50.00,95.00]-10 208kB ± 0% 8kB ± 0% -96.01% (p=0.000 n=9+10) RefreshRange/mixed-case/refresh_window=[75.00,99.00]-10 208kB ± 0% 8kB ± 0% -96.01% (p=0.000 n=10+10) RefreshRange/mixed-case/refresh_window=[0.00,95.00]-10 208kB ± 0% 8kB ± 0% -96.01% (p=0.000 n=10+10) RefreshRange/mixed-case/refresh_window=[50.00,99.00]-10 208kB ± 0% 8kB ± 0% -96.01% (p=0.000 n=10+10) RefreshRange/mixed-case/refresh_window=[75.00,95.00]-10 208kB ± 0% 8kB ± 0% -96.01% (p=0.000 n=10+10) RefreshRange/random-keys/refresh_window=[0.00,0.00]-10 29.3kB ± 4% 1.2kB ± 0% -95.95% (p=0.000 n=9+10) RefreshRange/mixed-case/refresh_window=[0.00,0.00]-10 9.07kB ±25% 1.13kB ± 0% -87.51% (p=0.000 n=10+10) name old allocs/op new allocs/op delta RefreshRange/linear-keys/refresh_window=[99.00,99.00]-10 18.8k ± 0% 0.0k ± 0% -99.91% (p=0.000 n=10+10) RefreshRange/linear-keys/refresh_window=[95.00,95.00]-10 18.1k ± 0% 0.0k ± 0% -99.91% (p=0.000 n=10+10) RefreshRange/mixed-case/refresh_window=[95.00,95.00]-10 17.5k ± 0% 0.0k ± 0% -99.90% (p=0.000 n=9+10) RefreshRange/linear-keys/refresh_window=[75.00,75.00]-10 14.4k ± 0% 0.0k ± 0% -99.88% (p=0.000 n=10+10) RefreshRange/linear-keys/refresh_window=[50.00,50.00]-10 9.81k ± 0% 0.02k ± 0% -99.83% (p=0.000 n=10+10) RefreshRange/linear-keys/refresh_window=[95.00,99.00]-10 18.2k ± 0% 0.1k ± 0% -99.69% (p=0.000 n=9+10) RefreshRange/linear-keys/refresh_window=[75.00,95.00]-10 14.4k ± 0% 0.1k ± 0% -99.61% (p=0.000 n=10+10) RefreshRange/linear-keys/refresh_window=[75.00,99.00]-10 14.4k ± 0% 0.1k ± 0% -99.61% (p=0.000 n=10+10) RefreshRange/mixed-case/refresh_window=[99.00,99.00]-10 18.2k ± 0% 0.1k ± 2% -99.51% (p=0.000 n=10+8) RefreshRange/linear-keys/refresh_window=[50.00,95.00]-10 9.62k ± 0% 0.06k ± 0% -99.40% (p=0.000 n=10+10) RefreshRange/linear-keys/refresh_window=[50.00,75.00]-10 9.62k ± 0% 0.06k ± 0% -99.40% (p=0.000 n=10+10) RefreshRange/linear-keys/refresh_window=[50.00,99.00]-10 9.62k ± 0% 0.06k ± 0% -99.40% (p=0.000 n=9+10) RefreshRange/mixed-case/refresh_window=[75.00,75.00]-10 13.8k ± 0% 0.2k ±28% -98.85% (p=0.000 n=9+10) RefreshRange/mixed-case/refresh_window=[50.00,50.00]-10 9.25k ± 0% 0.17k ± 6% -98.11% (p=0.000 n=10+9) RefreshRange/random-keys/refresh_window=[75.00,75.00]-10 14.2k ± 0% 0.3k ± 8% -97.89% (p=0.000 n=10+9) RefreshRange/linear-keys/refresh_window=[0.00,0.00]-10 535 ± 3% 17 ± 0% -96.82% (p=0.000 n=10+10) RefreshRange/random-keys/refresh_window=[99.00,99.00]-10 18.6k ± 0% 0.8k ± 3% -95.56% (p=0.000 n=10+8) RefreshRange/random-keys/refresh_window=[95.00,95.00]-10 17.9k ± 0% 0.8k ± 2% -95.34% (p=0.000 n=10+8) RefreshRange/random-keys/refresh_window=[0.00,0.00]-10 351 ± 3% 18 ± 0% -94.88% (p=0.000 n=9+10) RefreshRange/random-keys/refresh_window=[50.00,50.00]-10 9.59k ± 0% 0.95k ±12% -90.05% (p=0.000 n=9+10) RefreshRange/mixed-case/refresh_window=[0.00,0.00]-10 73.7 ±10% 17.0 ± 0% -76.93% (p=0.000 n=10+10) RefreshRange/linear-keys/refresh_window=[0.00,50.00]-10 80.0 ± 0% 56.0 ± 0% -30.00% (p=0.000 n=10+10) RefreshRange/random-keys/refresh_window=[0.00,50.00]-10 79.0 ± 0% 56.0 ± 0% -29.11% (p=0.000 n=9+10) RefreshRange/random-keys/refresh_window=[95.00,99.00]-10 79.0 ± 0% 56.0 ± 0% -29.11% (p=0.000 n=10+10) RefreshRange/mixed-case/refresh_window=[0.00,50.00]-10 79.0 ± 0% 56.0 ± 0% -29.11% (p=0.000 n=10+10) RefreshRange/mixed-case/refresh_window=[0.00,75.00]-10 79.0 ± 0% 56.0 ± 0% -29.11% (p=0.000 n=10+10) RefreshRange/mixed-case/refresh_window=[50.00,75.00]-10 79.0 ± 0% 56.0 ± 0% -29.11% (p=0.000 n=10+10) RefreshRange/mixed-case/refresh_window=[95.00,99.00]-10 79.0 ± 0% 56.0 ± 0% -29.11% (p=0.000 n=10+10) RefreshRange/linear-keys/refresh_window=[0.00,75.00]-10 80.0 ± 0% 58.0 ± 0% -27.50% (p=0.000 n=9+10) RefreshRange/linear-keys/refresh_window=[0.00,95.00]-10 80.0 ± 0% 58.0 ± 0% -27.50% (p=0.002 n=8+10) RefreshRange/linear-keys/refresh_window=[0.00,99.00]-10 79.7 ± 1% 58.0 ± 0% -27.23% (p=0.000 n=10+10) RefreshRange/random-keys/refresh_window=[0.00,75.00]-10 79.0 ± 0% 58.0 ± 0% -26.58% (p=0.000 n=10+10) RefreshRange/random-keys/refresh_window=[0.00,95.00]-10 79.0 ± 0% 58.0 ± 0% -26.58% (p=0.000 n=10+10) RefreshRange/random-keys/refresh_window=[0.00,99.00]-10 79.0 ± 0% 58.0 ± 0% -26.58% (p=0.000 n=10+10) RefreshRange/random-keys/refresh_window=[50.00,75.00]-10 79.0 ± 0% 58.0 ± 0% -26.58% (p=0.000 n=10+10) RefreshRange/random-keys/refresh_window=[50.00,95.00]-10 79.0 ± 0% 58.0 ± 0% -26.58% (p=0.000 n=10+10) RefreshRange/random-keys/refresh_window=[50.00,99.00]-10 79.0 ± 0% 58.0 ± 0% -26.58% (p=0.000 n=10+10) RefreshRange/random-keys/refresh_window=[75.00,95.00]-10 79.0 ± 0% 58.0 ± 0% -26.58% (p=0.000 n=10+10) RefreshRange/random-keys/refresh_window=[75.00,99.00]-10 79.0 ± 0% 58.0 ± 0% -26.58% (p=0.000 n=10+10) RefreshRange/mixed-case/refresh_window=[0.00,95.00]-10 79.0 ± 0% 60.0 ± 0% -24.05% (p=0.000 n=10+10) RefreshRange/mixed-case/refresh_window=[0.00,99.00]-10 79.0 ± 0% 60.0 ± 0% -24.05% (p=0.000 n=10+10) RefreshRange/mixed-case/refresh_window=[50.00,95.00]-10 79.0 ± 0% 60.0 ± 0% -24.05% (p=0.000 n=10+10) RefreshRange/mixed-case/refresh_window=[50.00,99.00]-10 79.0 ± 0% 60.0 ± 0% -24.05% (p=0.000 n=10+10) RefreshRange/mixed-case/refresh_window=[75.00,95.00]-10 79.0 ± 0% 60.0 ± 0% -24.05% (p=0.000 n=10+10) RefreshRange/mixed-case/refresh_window=[75.00,99.00]-10 79.0 ± 0% 60.0 ± 0% -24.05% (p=0.000 n=10+10) ``` ---- Release note (performance improvement): transaction read refresh operations performed during optimistic concurrency control's validation phase now use a time-bound file filter when scanning the LSM tree. This allows these operations to avoid scanning files that contain no keys written since the transaction originally performed its reads.
nvanbenschoten · Jan 27, 2022 · 57cc377 · 57cc377
1 parent a0ee732
commit 57cc377
Show file tree

Hide file tree

Showing 6 changed files with 400 additions and 41 deletions.
diff --git a/pkg/kv/kvserver/batcheval/.gitignore b/pkg/kv/kvserver/batcheval/.gitignore
@@ -0,0 +1 @@
+refresh_range_bench_data_*
diff --git a/pkg/kv/kvserver/batcheval/BUILD.bazel b/pkg/kv/kvserver/batcheval/BUILD.bazel
@@ -104,6 +104,7 @@ go_test(
         "cmd_lease_test.go",
         "cmd_query_resolved_timestamp_test.go",
         "cmd_recover_txn_test.go",
+        "cmd_refresh_range_bench_test.go",
         "cmd_refresh_range_test.go",
         "cmd_refresh_test.go",
         "cmd_resolve_intent_test.go",
@@ -149,6 +150,7 @@ go_test(
         "//pkg/testutils/sstutil",
         "//pkg/testutils/testcluster",
         "//pkg/util",
+        "//pkg/util/encoding",
         "//pkg/util/hlc",
         "//pkg/util/leaktest",
         "//pkg/util/log",
@@ -158,6 +160,9 @@ go_test(
         "//pkg/util/uint128",
         "//pkg/util/uuid",
         "@com_github_cockroachdb_errors//:errors",
+        "@com_github_cockroachdb_errors//oserror",
+        "@com_github_cockroachdb_pebble//:pebble",
+        "@com_github_cockroachdb_pebble//vfs",
         "@com_github_stretchr_testify//assert",
         "@com_github_stretchr_testify//require",
     ],

diff --git a/pkg/kv/kvserver/batcheval/cmd_refresh_range.go b/pkg/kv/kvserver/batcheval/cmd_refresh_range.go
@@ -15,11 +15,25 @@ import (
 
 	"github.com/cockroachdb/cockroach/pkg/kv/kvserver/batcheval/result"
 	"github.com/cockroachdb/cockroach/pkg/roachpb"
+	"github.com/cockroachdb/cockroach/pkg/settings"
 	"github.com/cockroachdb/cockroach/pkg/storage"
+	"github.com/cockroachdb/cockroach/pkg/storage/enginepb"
+	"github.com/cockroachdb/cockroach/pkg/util"
+	"github.com/cockroachdb/cockroach/pkg/util/hlc"
 	"github.com/cockroachdb/cockroach/pkg/util/log"
+	"github.com/cockroachdb/cockroach/pkg/util/protoutil"
+	"github.com/cockroachdb/cockroach/pkg/util/uuid"
 	"github.com/cockroachdb/errors"
 )
 
+// refreshRangeTBIEnabled controls whether we use a TBI during ranged refreshes.
+var refreshRangeTBIEnabled = settings.RegisterBoolSetting(
+	settings.SystemOnly,
+	"kv.refresh_range.time_bound_iterators.enabled",
+	"use time-bound iterators when performing ranged transaction refreshes",
+	util.ConstantWithMetamorphicTestBool("kv.refresh_range.time_bound_iterators.enabled", true),
+)
+
 func init() {
 	RegisterReadOnlyCommand(roachpb.RefreshRange, DefaultDeclareKeys, RefreshRange)
 }
@@ -50,40 +64,77 @@ func RefreshRange(
 		return result.Result{}, errors.AssertionFailedf("empty RefreshFrom: %s", args)
 	}
 
-	// Iterate over values until we discover any value written after the
-	// original timestamp, but before or at the current timestamp. Note that we
-	// iterate inconsistently, meaning that intents - including our own - are
-	// collected separately and the callback is only invoked on the latest
-	// committed version. Note also that we include tombstones, which must be
-	// considered as updates on refresh.
 	log.VEventf(ctx, 2, "refresh %s @[%s-%s]", args.Span(), refreshFrom, refreshTo)
-	intents, err := storage.MVCCIterate(
-		ctx, reader, args.Key, args.EndKey, refreshTo,
-		storage.MVCCScanOptions{
-			Inconsistent: true,
-			Tombstones:   true,
-		},
-		func(kv roachpb.KeyValue) error {
-			if ts := kv.Value.Timestamp; refreshFrom.Less(ts) {
-				return roachpb.NewRefreshFailedError(roachpb.RefreshFailedError_REASON_COMMITTED_VALUE, kv.Key, ts)
-			}
-			return nil
-		})
-	if err != nil {
-		return result.Result{}, err
-	}
+	tbi := refreshRangeTBIEnabled.Get(&cArgs.EvalCtx.ClusterSettings().SV)
+	return result.Result{}, refreshRange(reader, tbi, args.Span(), refreshFrom, refreshTo, h.Txn.ID)
+}
+
+// refreshRange iterates over the specified key span until it discovers a value
+// written after the refreshFrom timestamp but before or at the refreshTo
+// timestamp. The iteration observes MVCC tombstones, which must be considered
+// as conflicts during a refresh. The iteration also observes intents, and any
+// intent that is not owned by the specified txn ID is considered a conflict.
+//
+// If such a conflict is found, the function returns an error. Otherwise, no
+// error is returned.
+func refreshRange(
+	reader storage.Reader,
+	timeBoundIterator bool,
+	span roachpb.Span,
+	refreshFrom, refreshTo hlc.Timestamp,
+	txnID uuid.UUID,
+) error {
+	// Construct an incremental iterator with the desired time bounds. Incremental
+	// iterators will emit MVCC tombstones by default and will emit intents when
+	// configured to do so (see IntentPolicy).
+	iter := storage.NewMVCCIncrementalIterator(reader, storage.MVCCIncrementalIterOptions{
+		EnableTimeBoundIteratorOptimization: timeBoundIterator,
+		EndKey:                              span.EndKey,
+		StartTime:                           refreshFrom, // exclusive
+		EndTime:                             refreshTo,   // inclusive
+		IntentPolicy:                        storage.MVCCIncrementalIterIntentPolicyEmit,
+	})
+	defer iter.Close()
 
-	// Check if any intents which are not owned by this transaction were written
-	// at or beneath the refresh timestamp.
-	for _, i := range intents {
-		// Ignore our own intents.
-		if i.Txn.ID == h.Txn.ID {
-			continue
+	var meta enginepb.MVCCMetadata
+	iter.SeekGE(storage.MakeMVCCMetadataKey(span.Key))
+	for {
+		if ok, err := iter.Valid(); err != nil {
+			return err
+		} else if !ok {
+			break
 		}
-		// Return an error if an intent was written to the span.
-		return result.Result{}, roachpb.NewRefreshFailedError(roachpb.RefreshFailedError_REASON_INTENT,
-			i.Key, i.Txn.WriteTimestamp)
-	}
 
-	return result.Result{}, nil
+		key := iter.Key()
+		if !key.IsValue() {
+			// Found an intent. Check whether it is owned by this transaction.
+			// If so, proceed with iteration. Otherwise, return an error.
+			if err := protoutil.Unmarshal(iter.UnsafeValue(), &meta); err != nil {
+				return errors.Wrapf(err, "unmarshaling mvcc meta: %v", key)
+			}
+			if meta.IsInline() {
+				// Ignore inline MVCC metadata. We don't expect to see this in practice
+				// when performing a refresh of an MVCC keyspace.
+				iter.Next()
+				continue
+			}
+			if meta.Txn.ID == txnID {
+				// Ignore the transaction's own intent and skip past the corresponding
+				// provisional key-value. To do this, scan to the timestamp immediately
+				// before (i.e. the key immediately after) the provisional key.
+				iter.SeekGE(storage.MVCCKey{
+					Key:       key.Key,
+					Timestamp: meta.Timestamp.ToTimestamp().Prev(),
+				})
+				continue
+			}
+			return roachpb.NewRefreshFailedError(roachpb.RefreshFailedError_REASON_INTENT,
+				key.Key, meta.Txn.WriteTimestamp)
+		}
+
+		// If a committed value is found, return an error.
+		return roachpb.NewRefreshFailedError(roachpb.RefreshFailedError_REASON_COMMITTED_VALUE,
+			key.Key, key.Timestamp)
+	}
+	return nil
 }