@@ -51,13 +51,13 @@ if a transaction can be shown to always do the right thing when it is
5151run alone (before or after any other transaction), it will always do
5252the right thing in any mix of concurrent serializable transactions.
5353Where conflicts with other transactions would result in an
54- inconsistent state within the database, or an inconsistent view of
54+ inconsistent state within the database or an inconsistent view of
5555the data, a serializable transaction will block or roll back to
5656prevent the anomaly. The SQL standard provides a specific SQLSTATE
5757for errors generated when a transaction rolls back for this reason,
5858so that transactions can be retried automatically.
5959
60- Before version 9.1 PostgreSQL did not support a full serializable
60+ Before version 9.1, PostgreSQL did not support a full serializable
6161isolation level. A request for serializable transaction isolation
6262actually provided snapshot isolation. This has well known anomalies
6363which can allow data corruption or inconsistent views of the data
@@ -77,7 +77,7 @@ Serializable Isolation Implementation Strategies
7777
7878Techniques for implementing full serializable isolation have been
7979published and in use in many database products for decades. The
80- primary technique which has been used is Strict2 Phase Locking
80+ primary technique which has been used is StrictTwo- Phase Locking
8181(S2PL), which operates by blocking writes against data which has been
8282read by concurrent transactions and blocking any access (read or
8383write) against data which has been written by concurrent
@@ -112,54 +112,90 @@ visualize the difference between the serializable implementations
112112described above, is to consider that among transactions executing at
113113the serializable transaction isolation level, the results are
114114required to be consistent with some serial (one-at-a-time) execution
115- of the transactions[1]. How is that order determined in each?
115+ of the transactions
116116
117- S2PL locks rows used by the transaction in a way which blocks
118- conflicting access, so that at the moment of a successful commit it
119- is certain that no conflicting access has occurred. Some transactions
120- may have blocked, essentially being partially serialized with the
121- committing transaction, to allow this. Some transactions may have
122- been rolled back, due to cycles in the blocking. But with S2PL,
123- transactions can always be viewed as having occurred serially, in the
124- order of successful commit.
117+ In S2PL, each transaction locks any data it accesses. It holds the
118+ locks until committing, preventing other transactions from making
119+ conflicting accesses to the same data in the interim. Some
120+ transactions may have to be rolled back to prevent deadlock. But
121+ successful transactions can always be viewed as having occurred
122+ sequentially, in the order they committed.
125123
126124With snapshot isolation, reads never block writes, nor vice versa, so
127- there is much less actual serialization. The order in which
128- transactions appear to have executed is determined by something more
129- subtle than in S2PL: read/write dependencies. If a transaction
130- attempts to read data which is not visible to it because the
131- transaction which wrote it (or will later write it) is concurrent
132- (one of them was running when the other acquired its snapshot), then
133- the reading transaction appears to have executed first, regardless of
134- the actual sequence of transaction starts or commits (since it sees a
135- database state prior to that in which the other transaction leaves
136- it). If one transaction has both rw-dependencies in (meaning that a
137- concurrent transaction attempts to read data it writes) and out
138- (meaning it attempts to read data a concurrent transaction writes),
139- and a couple other conditions are met, there can appear to be a cycle
140- in execution order of the transactions. This is when the anomalies
141- occur.
142-
143- SSI works by watching for the conditions mentioned above, and rolling
144- back a transaction when needed to prevent any anomaly. The apparent
145- order of execution will always be consistent with any actual
146- serialization (i.e., a transaction which run by itself can always be
147- considered to have run after any transactions committed before it
148- started and before any transacton which starts after it commits); but
149- among concurrent transactions it will appear that the transaction on
150- the read side of a rw-dependency executed before the transaction on
151- the write side.
125+ more concurrency is possible. The order in which transactions appear
126+ to have executed is determined by something more subtle than in S2PL:
127+ read/write dependencies. If a transaction reads data, it appears to
128+ execute after the transaction that wrote the data it is reading.
129+ Similarly, if it updates data, it appears to execute after the
130+ transaction that wrote the previous version. These dependencies, which
131+ we call "wr-dependencies" and "ww-dependencies", are consistent with
132+ the commit order, because the first transaction must have committed
133+ before the second starts. However, there can also be dependencies
134+ between two *concurrent* transactions, i.e. where one was running when
135+ the other acquired its snapshot. These "rw-conflicts" occur when one
136+ transaction attempts to read data which is not visible to it because
137+ the transaction which wrote it (or will later write it) is
138+ concurrent. The reading transaction appears to have executed first,
139+ regardless of the actual sequence of transaction starts or commits,
140+ because it sees a database state prior to that in which the other
141+ transaction leaves it.
142+
143+ Anomalies occur when a cycle is created in the graph of dependencies:
144+ when a dependency or series of dependencies causes transaction A to
145+ appear to have executed before transaction B, but another series of
146+ dependencies causes B to appear before A. If that's the case, then
147+ the results can't be consistent with any serial execution of the
148+ transactions.
149+
150+
151+ SSI Algorithm
152+ -------------
153+
154+ Serializable transaction in PostgreSQL are implemented using
155+ Serializable Snapshot Isolation (SSI), based on the work of Cahill
156+ et al. Fundamentally, this allows snapshot isolation to run as it
157+ has, while monitoring for conditions which could create a serialization
158+ anomaly.
159+
160+ SSI is based on the observation [2] that each snapshot isolation
161+ anomaly corresponds to a cycle that contains a "dangerous structure"
162+ of two adjacent rw-conflict edges:
163+
164+ Tin ------> Tpivot ------> Tout
165+ rw rw
166+
167+ SSI works by watching for this dangerous structure, and rolling
168+ back a transaction when needed to prevent any anomaly. This means it
169+ only needs to track rw-conflicts between concurrent transactions, not
170+ wr- and ww-dependencies. It also means there is a risk of false
171+ positives, because not every dangerous structure corresponds to an
172+ actual serialization failure.
173+
174+ The PostgreSQL implementation uses two additional optimizations:
175+
176+ * Tout must commit before any other transaction in the cycle
177+ (see proof of Theorem 2.1 of [2]). We only roll back a transaction
178+ if Tout commits before Tpivot and Tin.
179+
180+ * if Tin is read-only, there can only be an anomaly if Tout committed
181+ before Tin takes its snapshot. This optimization is an original
182+ one. Proof:
183+
184+ - Because there is a cycle, there must be some transaction T0 that
185+ precedes Tin in the serial order. (T0 might be the same as Tout).
186+
187+ - The dependency between T0 and Tin can't be a rw-conflict,
188+ because Tin was read-only, so it must be a wr-dependency.
189+ Those can only occur if T0 committed before Tin started.
190+
191+ - Because Tout must commit before any other transaction in the
192+ cycle, it must commit before T0 commits -- and thus before Tin
193+ starts.
152194
153195
154196PostgreSQL Implementation
155197-------------------------
156198
157- The implementation of serializable transactions for PostgreSQL is
158- accomplished through Serializable Snapshot Isolation (SSI), based on
159- the work of Cahill, et al. Fundamentally, this allows snapshot
160- isolation to run as it has, while monitoring for conditions which
161- could create a serialization anomaly.
162-
163199 * Since this technique is based on Snapshot Isolation (SI), those
164200areas in PostgreSQL which don't use SI can't be brought under SSI.
165201This includes system tables, temporary tables, sequences, hint bit
@@ -180,7 +216,7 @@ lock or to use SELECT FOR SHARE or SELECT FOR UPDATE.
180216 * Those who want to continue to use snapshot isolation without
181217the additional protections of SSI (and the associated costs of
182218enforcing those protections), can use the REPEATABLE READ transaction
183- isolation level. This levelwill retain its legacy behavior, which
219+ isolation level. This levelretains its legacy behavior, which
184220is identical to the old SERIALIZABLE implementation and fully
185221consistent with the standard's requirements for the REPEATABLE READ
186222transaction isolation level.
@@ -236,7 +272,7 @@ in PostgreSQL, but tailored to the needs of SIREAD predicate locking,
236272are used. These refer to physical objects actually accessed in the
237273course of executing the query, to model the predicates through
238274inference. Anyone interested in this subject should review the
239- Hellerstein, Stonebraker and Hamilton paper[2 ], along with the
275+ Hellerstein, Stonebraker and Hamilton paper [3 ], along with the
240276locking papers referenced from that and the Cahill papers.
241277
242278Because the SIREAD locks don't block, traditional locking techniques
@@ -273,6 +309,15 @@ transaction already holds a write lock on any tuple representing the
273309row, since a rw-dependency would also create a ww-dependency which
274310has more aggressive enforcement and will thus prevent any anomaly.
275311
312+ * Modifying a heap tuple creates a rw-conflict with any transaction
313+ that holds a SIREAD lock on that tuple, or on the page or relation
314+ that contains it.
315+
316+ * Inserting a new tuple creates a rw-conflict with any transaction
317+ holding a SIREAD lock on the entire relation. It doesn't conflict with
318+ page-level locks, because page-level locks are only used to aggregate
319+ tuple locks. Unlike index page locks, they don't lock "gaps" on the page.
320+
276321
277322Index AM implementations
278323------------------------
@@ -296,13 +341,13 @@ need not generate a conflict, although an update which "moves" a row
296341into the scan must generate a conflict. While correctness allows
297342false positives, they should be minimized for performance reasons.
298343
299- Several optimizations are possible:
344+ Several optimizations are possible, though not all implemented yet :
300345
301346 * An index scan which is just finding the right position for an
302- index insertion or deletionneed not acquire a predicate lock.
347+ index insertion or deletionneeds not acquire a predicate lock.
303348
304349 * An index scan which is comparing for equality on the entire key
305- for a unique indexneed not acquire a predicate lock as long as a key
350+ for a unique indexneeds not acquire a predicate lock as long as a key
306351is found corresponding to a visible tuple which has not been modified
307352by another transaction -- there are no "between or around" gaps to
308353cover.
@@ -317,10 +362,10 @@ x = 1 AND x = 2), then no predicate lock is needed.
317362
318363Other index AM implementation considerations:
319364
320- *If a btree search discovers that no root page has yet been
321- created, a predicate lock on the indexrelation is required;
322- otherwise btree searches must get to the leaf level to determine
323- which tuples match, so predicate locks go there .
365+ *B-tree index searches acquire predicate locks only on the
366+ index *leaf* pages needed to lock theappropriate indexrange. If,
367+ however, a search discovers that no root page has yet been created, a
368+ predicate lock on the index relation is required .
324369
325370 * GiST searches can determine that there are no matches at any
326371level of the index, so there must be a predicate lock at each index
@@ -346,11 +391,6 @@ to be added from scratch.
346391
347392 2. The existing in-memory lock structures were not suitable for
348393tracking SIREAD locks.
349- * The database products used for the prototype
350- implementations for the papers used update-in-place with a rollback
351- log for their MVCC implementations, while PostgreSQL leaves the old
352- version of a row in place and adds a new tuple to represent the row
353- at a new location.
354394 * In PostgreSQL, tuple level locks are not held in RAM for
355395any length of time; lock information is written to the tuples
356396involved in the transactions.
@@ -450,18 +490,19 @@ there can't be a rw-conflict from T3 to T0.
450490
451491 o In both cases, we didn't need the T1 -> T3 edge.
452492
453- * Predicate locking in PostgreSQLwill start at the tuple level
454- when possible, with automatic conversion of multiple fine-grained
455- locks to coarserneed to avoid resource exhaustion.
456- The amount of memory used for these structureswill be configurable,
457- to balance RAM usage against SIREAD lock granularity.
493+ * Predicate locking in PostgreSQLstarts at the tuple level
494+ when possible. Multiple fine-grained locks are promoted to a single
495+ coarser- granularitylock asneeded to avoid resource exhaustion. The
496+ amount of memory used for these structuresis configurable, to balance
497+ RAM usage against SIREAD lock granularity.
458498
459- * A process-local copy of locks held by a process and the coarser
460- covering locks with counts, are kept to support granularity promotion
461- decisions with low CPU and locking overhead.
499+ * Each backend keeps a process-local table of the locks it holds.
500+ To support granularity promotion decisions with low CPU and locking
501+ overhead, this table also includes the coarser covering locks and the
502+ number of finer-granularity locks they cover.
462503
463- * Conflictswill be identified by looking for predicate locks
464- when tuples are written and looking at the MVCC information when
504+ * Conflictsare identified by looking for predicate locks
505+ when tuples are written, and by looking at the MVCC information when
465506tuples are read. There is no matching between two RAM-based locks.
466507
467508 * Because write locks are stored in the heap tuples rather than a
@@ -493,12 +534,12 @@ to be READ ONLY.)
493534 o We can more aggressively clean up conflicts, predicate
494535locks, and SSI transaction information.
495536
496- *Allow a READ ONLY transaction to "opt out" of SSI if there are
537+ *We allow a READ ONLY transaction to "opt out" of SSI if there are
497538no READ WRITE transactions which could cause the READ ONLY
498539transaction to ever become part of a "dangerous structure" of
499540overlapping transaction dependencies.
500541
501- *Allow the user to request that a READ ONLY transaction wait
542+ *We allow the user to request that a READ ONLY transaction wait
502543until the conditions are right for it to start in the "opt out" state
503544described above. We add a DEFERRABLE state to transactions, which is
504545specified and maintained in a way similar to READ ONLY. It is
@@ -538,28 +579,13 @@ address it?
538579replication solutions, like Postgres-R, Slony, pgpool, HS/SR, etc.
539580This is related to the "WAL file replay" issue.
540581
541- * Weak-memory-ordering machines. Make sure that shared memory
542- access which involves visibility across multiple transactions uses
543- locks as needed to avoid problems. On the other hand, ensure that we
544- really need volatile where we're using it.
545- http://archives.postgresql.org/pgsql-committers/2008-06/msg00228.php
546-
547582 * UNIQUE btree search for equality on all columns. Since a search
548583of a UNIQUE index using equality tests on all columns will lock the
549584heap tuple if an entry is found, it appears that there is no need to
550585get a predicate lock on the index in that case. A predicate lock is
551586still needed for such a search if a matching index entry which points
552587to a visible tuple is not found.
553588
554- * Planner index probes. To avoid problems with data skew at the
555- ends of an index which have historically caused bad plans, the
556- planner now probes the end of an index to see what the maximum or
557- minimum value is when a query appears to be requesting a range of
558- data outside what statistics shows is present. These planner checks
559- don't require predicate locking, but there's currently no easy way to
560- avoid it. What can we do to avoid predicate locking for such planner
561- activity?
562-
563589 * Minimize touching of shared memory. Should lists in shared
564590memory push entries which have just been returned to the front of the
565591available list, so they will be popped back off soon and some memory
@@ -573,13 +599,17 @@ Footnotes
573599[1] http://www.contrib.andrew.cmu.edu/~shadow/sql/sql1992.txt
574600Search for serial execution to find the relevant section.
575601
576- [2] http://db.cs.berkeley.edu/papers/fntdb07-architecture.pdf
577- Joseph M. Hellerstein, Michael Stonebraker and James Hamilton. 2007.
602+ [2] A. Fekete et al. Making Snapshot Isolation Serializable. In ACM
603+ Transactions on Database Systems 30:2, Jun. 2005.
604+ http://dx.doi.org/10.1145/1071610.1071615
605+
606+ [3] Joseph M. Hellerstein, Michael Stonebraker and James Hamilton. 2007.
578607Architecture of a Database System. Foundations and Trends(R) in
579608Databases Vol. 1, No. 2 (2007) 141-259.
609+ http://db.cs.berkeley.edu/papers/fntdb07-architecture.pdf
580610 Of particular interest:
581611 * 6.1 A Note on ACID
582612 * 6.2 A Brief Review of Serializability
583613 * 6.3 Locking and Latching
584614 * 6.3.1 Transaction Isolation Levels
585- * 6.5.3 Next-Key Locking: Physical Surrogates for Logical
615+ * 6.5.3 Next-Key Locking: Physical Surrogates for Logical Properties