Tom Lane wrote: > "Pavan Deolasee" <[EMAIL PROTECTED]> writes: >> Please see the version 14 of HOT patch attached. > > I expected to find either a large new README, or some pretty substantial > additions to existing README files, to document how this all works.
Here's an updated version of the README I posted earlier. It now reflects the changes to how pruning works. -- Heikki Linnakangas EnterpriseDB http://www.enterprisedb.com
Use case -------- The best use case for HOT is a table that's frequently UPDATEd, and is large enough that VACUUM is painful. On small tables that fit in cache, running VACUUM every few minutes isn't a problem. Heap-only tuples ---------------- When a HOT update is performed, the new tuple is placed on the same page as the old one, marked with the HEAP_ONLY_TUPLE flag. HEAP_ONLY_TUPLE means that there's no index pointers to the tuple, which allows pruning the chain in the future. The old tuple is marked with HEAP_HOT_UPDATE-flag, which means that the tuple pointed to by t_ctid is a heap-only tuple. That needs to be taken into account when vacuuming, so that we don't remove the root tuple in the update chain, when there's no index pointers to the later tuples. When doing an index scan, whenever we reach a non-visible tuple, we need to check if the tuple has been HOT-updated (== HEAP_HOT_UPDATE flag is set). If so, we need to follow the ctid pointer until we reach a visible one, or one that hasn't been HOT-updated. Sequential scans (and bitmap heap scans with a lossy bitmap) don't need to pay attention to the flags. The pre-requirements for doing a HOT update is that none of the indexed columns are changed. That's checked at execution time, comparing the binary representation of the old and new values. That means that dummy updates, like "UPDATE foo SET col1 = ?", where ? is the same as the old value can be HOT. In addition to the above, there needs to be room on the page for the new tuple. If the page is full, we try to make room by pruning the page. Pruning ------- Pruning is a lightweight vacuum operation that can be run on a single page, with no need to scan indexes, but it only removes dead HOT tuples. Other dead tuples are truncated, leaving only a redirected dead line pointer. The removed tuples are compacted away using PageRepairFragmentation, like in normal vacuum. There's two reasons to prune a page: to make room on the page for future updates, and to shorten HOT chains to make index lookups cheaper. When accessing a page with HOT updated tuples on it, and less than a certain threshold of free space, we try to prune it. To do that, we need to take a vacuum strength lock on the buffer. If that fails, we don't prune; the theory is that you usually do get the lock, and if you don't, you'll get to try again next time. It would be more logical to do the pruning in heap_update when the page is full, but by the time we get there we have already pinned the page and have references to tuples on it, so we can't start moving tuples around it. Also, that alone wouldn't address the desire to keep HOT chains short, to avoid overhead of traversing long chains on index lookups. To reclaim the index-visible (i.e. first) tuple in a HOT chain, the line pointer is turned into a redirecting line pointer that points to the line pointer of the next tuple in the chain. When the last live tuple in an update chain becomes dead (after a DELETE or a cold update), the redirecting line pointer is marked as redirected dead. That allows us to immediately reuse the space, sans the line pointer itself. We've effectively resurrected the "truncate dead tuples to just line pointer" idea that has been proposed and rejected before because of fear of line pointer bloat. To limit the damage in worst case, and to keep numerous arrays as well as the bitmaps in bitmap scans reasonably sized, the maximum number of line pointers (MaxHeapTuplesPerPage) is somewhat arbitrarily capped at 2 * what it was before. VACUUM FULL ----------- To make vacuum full work, any DEAD tuples in the middle of an update chain needs to be removed (see comments at the top of heap_prune_hotchain_hard for details). Vacuum full performs a more aggressive pruning that not only removes dead tuples at the beginning of an update chain, it scans the whole chain and removes any intermediate dead tuples as well. Vacuum ------ There's not much changes to regular vacuum. It removes dead HOT tuples, like pruning, and cleans up any redirected dead line pointers. In lazy vacuum, we must not freeze a tuple that's in the middle of an update chain. That can happen when a tuple has xmin > xmax; it's the same scenario that requires "hard pruning" in VACUUM FULL. Freezing such tuples will break the check that xmin and xmax matches when following the chain. It's not a problem without HOT, because the preceding tuple in the chain must be dead as well so no-one will try to follow the chain, but with HOT the preceding tuple would be DEAD_CHAIN, and someone might still need to follow the chain to find the live tuple. We avoid that by just not freezing such tuples. They can be frozen eventually, when the xmax of the preceding tuple is < OldestXmin as well. Statistics ---------- XXX: How do HOT-updates affect statistics? How often do we need to run autovacuum? CREATE INDEX CREATE INDEX CONCURRENTLY ------------------------- I'm not very familiar with how these, so I'll just shut up.. Glossary -------- Heap-only tuple A heap tuple with no index pointers. Marked with HEAP_ONLY_TUPLE flag. HOT-updated tuple An updated tuple, so that the next tuple in the chain is a heap-only tuple. Marked with HEAP_HOT_UPDATE flag. Redirecting line pointer A line pointer that points to another line pointer. lp_len is set to a magic value (ITEMID_REDIRECTED), and lp_off is the OffsetNumber of the line pointer it points to. Redirected dead line pointer A stub line pointer, that doesn't point to anything, but can't be removed or reused yet because there is index pointers to it. Semantically same as a dead tuple. Root tuple The first tuple in a HOT update chain, that indexes point to. Update chain A chain of updated tuples, so that each tuple's ctid points to the next tuple in the chain. A HOT update chain is an update chain that consists of a root tuple and one or more heap-only tuples. An update chain can contain both HOT and non-HOT (cold) updated tuples. Cold update A normal, non-HOT update. HOT update An UPDATE, where the new tuple becomes a heap-only-tuple, and no index entries are made. DEAD_CHAIN (HEAPTUPLE_DEAD_CHAIN) New return value for HeapTupleSatisfiesVacuum, which means that the tuple is not visible to anyone, but it's been HOT updated so we can't remove it yet because the following tuples in the chain would become inaccessible from indexes.
---------------------------(end of broadcast)--------------------------- TIP 9: In versions below 8.0, the planner will ignore your desire to choose an index scan if your joining column's datatypes do not match