diff options
Diffstat (limited to 'src/third_party/wiredtiger/src/include/reconcile.h')
| -rw-r--r-- | src/third_party/wiredtiger/src/include/reconcile.h | 89 |
1 files changed, 37 insertions, 52 deletions
diff --git a/src/third_party/wiredtiger/src/include/reconcile.h b/src/third_party/wiredtiger/src/include/reconcile.h index 22f63ae4ff4..1c02f8353c6 100644 --- a/src/third_party/wiredtiger/src/include/reconcile.h +++ b/src/third_party/wiredtiger/src/include/reconcile.h @@ -33,12 +33,9 @@ struct __wt_reconcile { /* Track the page's min/maximum transactions. */ uint64_t max_txn; - wt_timestamp_t max_timestamp; - - /* Lookaside boundary tracking. */ - uint64_t unstable_txn; - wt_timestamp_t unstable_durable_timestamp; - wt_timestamp_t unstable_timestamp; + wt_timestamp_t max_ts; + wt_timestamp_t max_ondisk_ts; + wt_timestamp_t min_skipped_ts; u_int updates_seen; /* Count of updates seen. */ u_int updates_unstable; /* Count of updates not visible_all. */ @@ -56,43 +53,36 @@ struct __wt_reconcile { bool leave_dirty; /* - * Track if reconciliation has seen any overflow items. If a leaf page - * with no overflow items is written, the parent page's address cell is - * set to the leaf-no-overflow type. This means we can delete the leaf - * page without reading it because we don't have to discard any overflow + * Track if reconciliation has seen any overflow items. If a leaf page with no overflow items is + * written, the parent page's address cell is set to the leaf-no-overflow type. This means we + * can delete the leaf page without reading it because we don't have to discard any overflow * items it might reference. * - * The test test is per-page reconciliation, that is, once we see an - * overflow item on the page, all subsequent leaf pages written for the - * page will not be leaf-no-overflow type, regardless of whether or not - * they contain overflow items. In other words, leaf-no-overflow is not - * guaranteed to be set on every page that doesn't contain an overflow - * item, only that if it is set, the page contains no overflow items. - * XXX - * This was originally done because raw compression couldn't do better, - * now that raw compression has been removed, we should do better. + * The test test is per-page reconciliation, that is, once we see an overflow item on the page, + * all subsequent leaf pages written for the page will not be leaf-no-overflow type, regardless + * of whether or not they contain overflow items. In other words, leaf-no-overflow is not + * guaranteed to be set on every page that doesn't contain an overflow item, only that if it is + * set, the page contains no overflow items. XXX This was originally done because raw + * compression couldn't do better, now that raw compression has been removed, we should do + * better. */ bool ovfl_items; /* - * Track if reconciliation of a row-store leaf page has seen empty (zero - * length) values. We don't write out anything for empty values, so if - * there are empty values on a page, we have to make two passes over the - * page when it's read to figure out how many keys it has, expensive in - * the common case of no empty values and (entries / 2) keys. Likewise, - * a page with only empty values is another common data set, and keys on - * that page will be equal to the number of entries. In both cases, set - * a flag in the page's on-disk header. + * Track if reconciliation of a row-store leaf page has seen empty (zero length) values. We + * don't write out anything for empty values, so if there are empty values on a page, we have to + * make two passes over the page when it's read to figure out how many keys it has, expensive in + * the common case of no empty values and (entries / 2) keys. Likewise, a page with only empty + * values is another common data set, and keys on that page will be equal to the number of + * entries. In both cases, set a flag in the page's on-disk header. * - * The test is per-page reconciliation as described above for the - * overflow-item test. + * The test is per-page reconciliation as described above for the overflow-item test. */ bool all_empty_value, any_empty_value; /* - * Reconciliation gets tricky if we have to split a page, which happens - * when the disk image we create exceeds the page type's maximum disk - * image size. + * Reconciliation gets tricky if we have to split a page, which happens when the disk image we + * create exceeds the page type's maximum disk image size. * * First, the target size of the page we're building. */ @@ -106,31 +96,26 @@ struct __wt_reconcile { uint32_t min_split_size; /* Minimum split page size */ /* - * We maintain two split chunks in the memory during reconciliation to - * be written out as pages. As we get to the end of the data, if the - * last one turns out to be smaller than the minimum split size, we go - * back into the penultimate chunk and split at this minimum split size - * boundary. This moves some data from the penultimate chunk to the last - * chunk, hence increasing the size of the last page written without - * decreasing the penultimate page size beyond the minimum split size. - * For this reason, we maintain an expected split percentage boundary - * and a minimum split percentage boundary. + * We maintain two split chunks in the memory during reconciliation to be written out as pages. + * As we get to the end of the data, if the last one turns out to be smaller than the minimum + * split size, we go back into the penultimate chunk and split at this minimum split size + * boundary. This moves some data from the penultimate chunk to the last chunk, hence increasing + * the size of the last page written without decreasing the penultimate page size beyond the + * minimum split size. For this reason, we maintain an expected split percentage boundary and a + * minimum split percentage boundary. * - * Chunks are referenced by current and previous pointers. In case of a - * split, previous references the first chunk and current switches to - * the second chunk. If reconciliation generates more split chunks, the - * the previous chunk is written to the disk and current and previous - * swap. + * Chunks are referenced by current and previous pointers. In case of a split, previous + * references the first chunk and current switches to the second chunk. If reconciliation + * generates more split chunks, the previous chunk is written to the disk and current and + * previous swap. */ struct __wt_rec_chunk { /* - * The recno and entries fields are the starting record number - * of the split chunk (for column-store splits), and the number - * of entries in the split chunk. + * The recno and entries fields are the starting record number of the split chunk (for + * column-store splits), and the number of entries in the split chunk. * - * The key for a row-store page; no column-store key is needed - * because the page's recno, stored in the recno field, is the - * column-store key. + * The key for a row-store page; no column-store key is needed because the page's recno, + * stored in the recno field, is the column-store key. */ uint32_t entries; uint64_t recno; |