Add a few entries to the tail of time mapping, to see old values.
authorKevin Grittner <[email protected]>
Fri, 29 Apr 2016 21:46:08 +0000 (16:46 -0500)
committerKevin Grittner <[email protected]>
Fri, 29 Apr 2016 21:46:08 +0000 (16:46 -0500)
Without a few entries beyond old_snapshot_threshold, the lookup
would often fail, resulting in the more aggressive pruning or
vacuum being skipped often enough to matter.  This was very clearly
shown by a python test script posted by Ants Aasma, and was likely
a factor in an earlier but somewhat less clear-cut test case posted
by Jeff Janes.

This patch makes no change to the logic, per se -- it just makes
the array of mapping entries big enough to make lookup misses based
on timing much less likely.  An occasional miss is still possible
if a thread stalls for more than 10 minutes, but that does not
create any problem with correctness of behavior.  Besides, if
things are so busy that a thread is stalling for more than 10
minutes, it is probably OK to skip the more aggressive cleanup at
that particular point in time.

src/backend/utils/time/snapmgr.c
src/include/utils/snapmgr.h

index 35f338133aa532a1ebe0221bedef935479ba048f..0a9a231f59708026fb351ebe272367d6b324a300 100644 (file)
@@ -92,10 +92,10 @@ typedef struct OldSnapshotControlData
     * Use a circular buffer with a head offset, a count of entries currently
     * used, and a timestamp corresponding to the xid at the head offset.  A
     * count_used value of zero means that there are no times stored; a
-    * count_used value of old_snapshot_threshold means that the buffer is
-    * full and the head must be advanced to add new entries.  Use timestamps
-    * aligned to minute boundaries, since that seems less surprising than
-    * aligning based on the first usage timestamp.
+    * count_used value of OLD_SNAPSHOT_TIME_MAP_ENTRIES means that the buffer
+    * is full and the head must be advanced to add new entries.  Use
+    * timestamps aligned to minute boundaries, since that seems less
+    * surprising than aligning based on the first usage timestamp.
     *
     * It is OK if the xid for a given time slot is from earlier than
     * calculated by adding the number of minutes corresponding to the
@@ -243,7 +243,7 @@ SnapMgrShmemSize(void)
    size = offsetof(OldSnapshotControlData, xid_by_minute);
    if (old_snapshot_threshold > 0)
        size = add_size(size, mul_size(sizeof(TransactionId),
-                                      old_snapshot_threshold));
+                                      OLD_SNAPSHOT_TIME_MAP_ENTRIES));
 
    return size;
 }
@@ -1643,7 +1643,7 @@ TransactionIdLimitedForOldSnapshots(TransactionId recentXmin,
                if (offset > oldSnapshotControl->count_used - 1)
                    offset = oldSnapshotControl->count_used - 1;
                offset = (oldSnapshotControl->head_offset + offset)
-                       % old_snapshot_threshold;
+                       % OLD_SNAPSHOT_TIME_MAP_ENTRIES;
                xlimit = oldSnapshotControl->xid_by_minute[offset];
 
                if (NormalTransactionIdFollows(xlimit, recentXmin))
@@ -1720,10 +1720,10 @@ MaintainOldSnapshotTimeMapping(int64 whenTaken, TransactionId xmin)
    LWLockAcquire(OldSnapshotTimeMapLock, LW_EXCLUSIVE);
 
    Assert(oldSnapshotControl->head_offset >= 0);
-   Assert(oldSnapshotControl->head_offset < old_snapshot_threshold);
+   Assert(oldSnapshotControl->head_offset < OLD_SNAPSHOT_TIME_MAP_ENTRIES);
    Assert((oldSnapshotControl->head_timestamp % USECS_PER_MINUTE) == 0);
    Assert(oldSnapshotControl->count_used >= 0);
-   Assert(oldSnapshotControl->count_used <= old_snapshot_threshold);
+   Assert(oldSnapshotControl->count_used <= OLD_SNAPSHOT_TIME_MAP_ENTRIES);
 
    if (oldSnapshotControl->count_used == 0)
    {
@@ -1750,7 +1750,7 @@ MaintainOldSnapshotTimeMapping(int64 whenTaken, TransactionId xmin)
        int     bucket = (oldSnapshotControl->head_offset
                          + ((ts - oldSnapshotControl->head_timestamp)
                             / USECS_PER_MINUTE))
-                        % old_snapshot_threshold;
+                        % OLD_SNAPSHOT_TIME_MAP_ENTRIES;
 
        if (TransactionIdPrecedes(oldSnapshotControl->xid_by_minute[bucket], xmin))
            oldSnapshotControl->xid_by_minute[bucket] = xmin;
@@ -1763,7 +1763,7 @@ MaintainOldSnapshotTimeMapping(int64 whenTaken, TransactionId xmin)
 
        oldSnapshotControl->head_timestamp = ts;
 
-       if (advance >= old_snapshot_threshold)
+       if (advance >= OLD_SNAPSHOT_TIME_MAP_ENTRIES)
        {
            /* Advance is so far that all old data is junk; start over. */
            oldSnapshotControl->head_offset = 0;
@@ -1777,12 +1777,12 @@ MaintainOldSnapshotTimeMapping(int64 whenTaken, TransactionId xmin)
 
            for (i = 0; i < advance; i++)
            {
-               if (oldSnapshotControl->count_used == old_snapshot_threshold)
+               if (oldSnapshotControl->count_used == OLD_SNAPSHOT_TIME_MAP_ENTRIES)
                {
                    /* Map full and new value replaces old head. */
                    int     old_head = oldSnapshotControl->head_offset;
 
-                   if (old_head == (old_snapshot_threshold - 1))
+                   if (old_head == (OLD_SNAPSHOT_TIME_MAP_ENTRIES - 1))
                        oldSnapshotControl->head_offset = 0;
                    else
                        oldSnapshotControl->head_offset = old_head + 1;
@@ -1793,7 +1793,7 @@ MaintainOldSnapshotTimeMapping(int64 whenTaken, TransactionId xmin)
                    /* Extend map to unused entry. */
                    int     new_tail = (oldSnapshotControl->head_offset
                                        + oldSnapshotControl->count_used)
-                                      % old_snapshot_threshold;
+                                      % OLD_SNAPSHOT_TIME_MAP_ENTRIES;
 
                    oldSnapshotControl->count_used++;
                    oldSnapshotControl->xid_by_minute[new_tail] = xmin;
index 6a80f0ba5508d09586e8238d060c19a447ba4e28..42706966f10b10ba9aa75676dc51de011174c8ee 100644 (file)
 #include "utils/snapshot.h"
 
 
+/*
+ * The structure used to map times to TransactionId values for the "snapshot
+ * too old" feature must have a few entries at the tail to hold old values;
+ * otherwise the lookup will often fail and the expected early pruning or
+ * vacuum will not usually occur.  It is best if this padding is for a number
+ * of minutes greater than a thread would normally be stalled, but it's OK if
+ * early vacuum opportunities are occasionally missed, so there's no need to
+ * use an extreme value or get too fancy.  10 minutes seems plenty.
+ */
+#define OLD_SNAPSHOT_PADDING_ENTRIES 10
+#define OLD_SNAPSHOT_TIME_MAP_ENTRIES (old_snapshot_threshold + OLD_SNAPSHOT_PADDING_ENTRIES)
+
+
 /* GUC variables */
 extern PGDLLIMPORT int old_snapshot_threshold;