Merge pull request #792 from wiredtiger/compact-lsm

Allow LSM trees to be compacted.

18 files changed, 249 insertions, 83 deletions

diff --git a/bench/wtperf/runners/medium-lsm-sleep.wtperf b/bench/wtperf/runners/medium-lsm-compact.wtperf
index bd734e06e55..5393cdbfeba 100644
--- a/bench/wtperf/runners/medium-lsm-sleep.wtperf
+++ b/bench/wtperf/runners/medium-lsm-compact.wtperf
@@ -1,9 +1,9 @@
 # wtperf options file: medium lsm configuration
 conn_config="cache_size=1G"
 table_config="lsm=(chunk_size=100MB,merge_threads=2),type=lsm"
-merge_sleep=-1
 icount=50000000
-report_interval=5
-run_time=120
 populate_threads=1
+compact=true
 threads=((count=16,read=1))
+run_time=120
+report_interval=5
diff --git a/bench/wtperf/wtperf.c b/bench/wtperf/wtperf.c
index 271a43e1aad..f817f0eff1d 100644
--- a/bench/wtperf/wtperf.c
+++ b/bench/wtperf/wtperf.c
@@ -742,13 +742,14 @@ err:		g_error = g_stop = 1;
 static int
 execute_populate(CONFIG *cfg)
 {
+	WT_SESSION *session;
 	struct timespec start, stop;
 	double secs;
 	uint64_t last_ops;
 	uint32_t interval;
-	u_int sleepsec;
 	int elapsed, ret;
 
+	session = NULL;
 	lprintf(cfg, 0, 1,
 	    "Starting %" PRIu32 " populate thread(s)", cfg->populate_threads);
 
@@ -781,8 +782,10 @@ execute_populate(CONFIG *cfg)
 		interval = 0;
 		g_insert_ops = sum_pop_ops(cfg);
 		lprintf(cfg, 0, 1,
-		    "%" PRIu64 " populate inserts in %" PRIu32 " secs",
-		    g_insert_ops - last_ops, cfg->report_interval);
+		    "%" PRIu64 " populate inserts (%" PRIu64 " of %"
+		    PRIu32 ") in %" PRIu32 " secs",
+		    g_insert_ops - last_ops, g_insert_ops,
+		    cfg->icount, cfg->report_interval);
 		last_ops = g_insert_ops;
 	}
 	if ((ret = __wt_epoch(NULL, &stop)) != 0) {
@@ -810,17 +813,20 @@ execute_populate(CONFIG *cfg)
 	    "Load time: %.2f\n" "load ops/sec: %.2f", secs, cfg->icount / secs);
 
 	/*
-	 * If configured, sleep for awhile to allow LSM merging to complete in
-	 * the background.  If user specifies -1, then sleep for as long as it
-	 * took to load.
+	 * If configured, compact to allow LSM merging to complete.
 	 */
-	if (cfg->merge_sleep) {
-		if (cfg->merge_sleep < 0)
-			sleepsec = (u_int)(stop.tv_sec - start.tv_sec);
-		else
-			sleepsec = (u_int)cfg->merge_sleep;
-		lprintf(cfg, 0, 1, "Sleep %u seconds for merging", sleepsec);
-		(void)sleep(sleepsec);
+	if (cfg->compact) {
+		if ((ret = cfg->conn->open_session(
+		    cfg->conn, NULL, NULL, &session)) != 0) {
+			lprintf(cfg, ret, 0,
+			     "execute_populate: WT_CONNECTION.open_session");
+			return (ret);
+		}
+		lprintf(cfg, 0, 1, "Compact after populate");
+		ret = session->compact(session, cfg->uri, NULL);
+		session->close(session, NULL);
+		if (ret != 0)
+			return (ret);
 	}
 
 	/*
diff --git a/bench/wtperf/wtperf_opt.i b/bench/wtperf/wtperf_opt.i
index da5a63400ef..f2c8eb6a02f 100644
--- a/bench/wtperf/wtperf_opt.i
+++ b/bench/wtperf/wtperf_opt.i
@@ -81,6 +81,7 @@ DEF_OPT_AS_UINT32(checkpoint_interval, 120, "checkpoint every interval seconds")
 DEF_OPT_AS_UINT32(checkpoint_threads, 0, "number of checkpoint threads")
 DEF_OPT_AS_CONFIG_STRING(conn_config, "create",
     "connection configuration string")
+DEF_OPT_AS_BOOL(compact, 0, "post-populate compact for LSM merging activity")
 DEF_OPT_AS_BOOL(create, 1,
     "do population phase; false to use existing database")
 DEF_OPT_AS_UINT32(value_sz, 100, "value size")
@@ -88,8 +89,6 @@ DEF_OPT_AS_UINT32(icount, 5000, "number of records to initially populate")
 DEF_OPT_AS_BOOL(insert_rmw, 0,
     "execute a read prior to each insert in workload phase")
 DEF_OPT_AS_UINT32(key_sz, 20, "key size")
-DEF_OPT_AS_INT(merge_sleep, 0,
-    "post-populate sleep seconds for LSM merging activity (-1 for load time)")
 DEF_OPT_AS_BOOL(pareto, 0, "use pareto 80/20 distribution for random numbers")
 DEF_OPT_AS_UINT32(populate_ops_per_txn, 0,
     "number of operations to group into each transaction in the populate "
diff --git a/dist/api_data.py b/dist/api_data.py
index 3e04f4be3ef..8dee608a9f4 100644
--- a/dist/api_data.py
+++ b/dist/api_data.py
@@ -388,7 +388,13 @@ methods = {
 
 'session.close' : Method([]),
 
-'session.compact' : Method([]),
+'session.compact' : Method([
+	Config('timeout', '1200', r'''
+	    maximum amount of time to allow for compact in seconds. The
+		actual amount of time spent in compact may exceed the configured
+		value. A value of zero disables the timeout''',
+		type='int'),
+]),
 
 'session.create' :
 	    Method(table_only_meta + file_config + lsm_config + source_meta + [
diff --git a/src/config/config_def.c b/src/config/config_def.c
index 1ea0a9e4d71..056f2218af1 100644
--- a/src/config/config_def.c
+++ b/src/config/config_def.c
@@ -115,6 +115,11 @@ static const WT_CONFIG_CHECK confchk_session_checkpoint[] = {
 	{ NULL, NULL, NULL, NULL }
 };
 
+static const WT_CONFIG_CHECK confchk_session_compact[] = {
+	{ "timeout", "int", NULL, NULL},
+	{ NULL, NULL, NULL, NULL }
+};
+
 static const WT_CONFIG_CHECK confchk_lsm_subconfigs[] = {
 	{ "auto_throttle", "boolean", NULL, NULL },
 	{ "bloom", "boolean", NULL, NULL },
@@ -355,8 +360,8 @@ static const WT_CONFIG_ENTRY config_entries[] = {
 	  NULL
 	},
 	{ "session.compact",
-	  "",
-	  NULL
+	  "timeout=1200",
+	  confchk_session_compact
 	},
 	{ "session.create",
 	  "allocation_size=4KB,block_allocation=best,block_compressor=,"
diff --git a/src/docs/compact.dox b/src/docs/compact.dox
index 634bb98d9b5..c68415b5324 100644
--- a/src/docs/compact.dox
+++ b/src/docs/compact.dox
@@ -1,13 +1,13 @@
 /*! @page compaction Compaction
 
-The WT_SESSION::compact method can be used to compact a row- or column-store
-btree (log-structured merge trees cannot be compacted).
+The WT_SESSION::compact method can be used to compact btree and log-structured
+merge tree data sources.
 
-The data source does not have to be quiescent, compaction may be performed on
-a live data source.
-
-Because checkpoints named by the application are not discarded until
-explicitly removed or replaced, they may prevent WT_SESSION::compact
-from accomplishing anything.
+The data source does not have to be quiescent, compaction may be performed on a
+live data source.
 
+Because checkpoints named by the application are not discarded until explicitly
+removed or replaced, they may prevent WT_SESSION::compact from accomplishing
+anything.
+ *
  */
diff --git a/src/include/compact.h b/src/include/compact.h
new file mode 100644
index 00000000000..f799b75ebca
--- /dev/null
+++ b/src/include/compact.h
@@ -0,0 +1,12 @@
+/*-
+ * Copyright (c) 2008-2013 WiredTiger, Inc.
+ *	All rights reserved.
+ *
+ * See the file LICENSE for redistribution information.
+ */
+
+struct __wt_compact {
+	uint32_t	lsm_count;	/* Number of LSM trees seen */
+	uint32_t	file_count;	/* Number of files seen */
+	uint64_t	max_time;	/* Configured timeout */
+};
diff --git a/src/include/extern.h b/src/include/extern.h
index 4374a80dc2c..34cc0ae382b 100644
--- a/src/include/extern.h
+++ b/src/include/extern.h
@@ -899,6 +899,7 @@ extern int __wt_lsm_tree_lock( WT_SESSION_IMPL *session,
     int exclusive);
 extern int __wt_lsm_tree_unlock( WT_SESSION_IMPL *session,
     WT_LSM_TREE *lsm_tree);
+extern int __wt_lsm_compact(WT_SESSION_IMPL *session, const char *name);
 extern int __wt_lsm_tree_worker(WT_SESSION_IMPL *session,
     const char *uri,
     int (*file_func)(WT_SESSION_IMPL *,
@@ -1280,6 +1281,7 @@ extern int __wt_open_session(WT_CONNECTION_IMPL *conn,
     WT_EVENT_HANDLER *event_handler,
     const char *config,
     WT_SESSION_IMPL **sessionp);
+extern int __wt_compact_uri_analyze(WT_SESSION_IMPL *session, const char *uri);
 extern int __wt_session_compact( WT_SESSION *wt_session,
     const char *uri,
     const char *config);
diff --git a/src/include/lsm.h b/src/include/lsm.h
index cd957c9d96f..9000e4cb226 100644
--- a/src/include/lsm.h
+++ b/src/include/lsm.h
@@ -91,6 +91,7 @@ struct __wt_lsm_tree {
 
 	long throttle_sleep;		/* Rate limiting */
 	uint64_t chunk_fill_ms;		/* Estimate of time to fill a chunk */
+	uint64_t merge_progressing;	/* Bumped when merges are active */
 
 	/* Configuration parameters */
 	uint32_t bloom_bit_count;
@@ -123,10 +124,11 @@ struct __wt_lsm_tree {
 	size_t old_alloc;		/* Space allocated for old chunks */
 	u_int nold_chunks;		/* Number of old chunks */
 
-#define	WT_LSM_TREE_NEED_SWITCH	0x01
-#define	WT_LSM_TREE_OPEN	0x02
-#define	WT_LSM_TREE_THROTTLE	0x04
-#define	WT_LSM_TREE_WORKING	0x08
+#define	WT_LSM_TREE_COMPACTING	0x01
+#define	WT_LSM_TREE_NEED_SWITCH	0x02
+#define	WT_LSM_TREE_OPEN	0x04
+#define	WT_LSM_TREE_THROTTLE	0x08
+#define	WT_LSM_TREE_WORKING	0x10
 	uint32_t flags;
 };
 
diff --git a/src/include/session.h b/src/include/session.h
index 5c31280f979..a8c215b0e9c 100644
--- a/src/include/session.h
+++ b/src/include/session.h
@@ -70,6 +70,7 @@ struct __wt_session_impl {
 	TAILQ_HEAD(__cursors, __wt_cursor) cursors;
 
 	WT_CURSOR_BACKUP *bkp_cursor;	/* Hot backup cursor */
+	WT_COMPACT	 *compact;	/* Compact state */
 
 	WT_BTREE *metafile;		/* Metadata file */
 	void	*meta_track;		/* Metadata operation tracking */
diff --git a/src/include/wiredtiger.in b/src/include/wiredtiger.in
index 8d68b3240ad..0a1c7f3156c 100644
--- a/src/include/wiredtiger.in
+++ b/src/include/wiredtiger.in
@@ -901,7 +901,12 @@ struct __wt_session {
 	 * @param session the session handle
 	 * @param name the URI of the object to compact, such as
 	 * \c "table:stock"
-	 * @configempty{session.compact, see dist/api_data.py}
+	 * @configstart{session.compact, see dist/api_data.py}
+	 * @config{timeout, maximum amount of time to allow for compact in
+	 * seconds.  The actual amount of time spent in compact may exceed the
+	 * configured value.  A value of zero disables the timeout., an integer;
+	 * default \c 1200.}
+	 * @configend
 	 * @errors
 	 */
 	int __F(compact)(WT_SESSION *session,
diff --git a/src/include/wt_internal.h b/src/include/wt_internal.h
index d36656be947..3377f4d721e 100644
--- a/src/include/wt_internal.h
+++ b/src/include/wt_internal.h
@@ -83,6 +83,8 @@ struct __wt_col_rle;
     typedef struct __wt_col_rle WT_COL_RLE;
 struct __wt_colgroup;
     typedef struct __wt_colgroup WT_COLGROUP;
+struct __wt_compact;
+    typedef struct __wt_compact WT_COMPACT;
 struct __wt_condvar;
     typedef struct __wt_condvar WT_CONDVAR;
 struct __wt_config;
@@ -237,6 +239,7 @@ struct __wt_update;
 #include "btree.h"
 #include "cache.h"
 #include "config.h"
+#include "compact.h"
 #include "cursor.h"
 #include "dlh.h"
 #include "error.h"
diff --git a/src/lsm/lsm_merge.c b/src/lsm/lsm_merge.c
index eebcf02c93f..f92b8610f80 100644
--- a/src/lsm/lsm_merge.c
+++ b/src/lsm/lsm_merge.c
@@ -61,6 +61,7 @@ __wt_lsm_merge(
 	uint32_t generation, start_id;
 	uint64_t insert_count, record_count, chunk_size;
 	u_int dest_id, end_chunk, i, merge_min, nchunks, start_chunk;
+	u_int max_generation_gap;
 	int create_bloom;
 	const char *cfg[3];
 
@@ -71,13 +72,15 @@ __wt_lsm_merge(
 	start_id = 0;
 
 	/*
-	 * If the tree is open read-only, be very aggressive.  Otherwise, we
-	 * can spend a long time waiting for merges to start in read-only
-	 * applications.
+	 * If the tree is open read-only or we are compacting, be very
+	 * aggressive.  Otherwise, we can spend a long time waiting for merges
+	 * to start in read-only applications.
 	 */
-	if (!lsm_tree->modified)
+	if (!lsm_tree->modified ||
+	    F_ISSET(lsm_tree, WT_LSM_TREE_COMPACTING))
 		aggressive = 10;
 	merge_min = (aggressive > 5) ? 2 : lsm_tree->merge_min;
+	max_generation_gap = 1 + aggressive / 5;
 
 	/*
 	 * If there aren't any chunks to merge, or some of the chunks aren't
@@ -153,13 +156,16 @@ __wt_lsm_merge(
 			break;
 
 		/*
-		 * If we have enough chunks for a merge and the next chunk is
-		 * in a different generation, stop.
+		 * In normal operation, if we have enough chunks for a merge
+		 * and the next chunk is in a different generation, stop.
+		 * In aggressive mode, look for the biggest merge we can do.
 		 */
 		if (nchunks >= merge_min) {
 			previous = lsm_tree->chunk[start_chunk];
-			if (chunk->generation > previous->generation &&
-			    previous->generation <= youngest->generation + 1)
+			if (previous->generation <=
+				youngest->generation + max_generation_gap &&
+			    chunk->generation >
+				previous->generation + max_generation_gap - 1)
 				break;
 		}
 
@@ -188,7 +194,8 @@ __wt_lsm_merge(
 		 * generations.
 		 */
 		if (nchunks < merge_min ||
-		    chunk->generation > youngest->generation + 1) {
+		    chunk->generation >
+		    youngest->generation + max_generation_gap) {
 			for (i = 0; i < nchunks; i++)
 				F_CLR(lsm_tree->chunk[start_chunk + i],
 				    WT_LSM_CHUNK_MERGING);
@@ -262,6 +269,7 @@ __wt_lsm_merge(
 				WT_ERR(EINTR);
 			WT_STAT_FAST_CONN_INCRV(session,
 			    lsm_rows_merged, LSM_MERGE_CHECK_INTERVAL);
+			++lsm_tree->merge_progressing;
 		}
 
 		WT_ERR(src->get_key(src, &key));
@@ -276,6 +284,7 @@ __wt_lsm_merge(
 
 	WT_STAT_FAST_CONN_INCRV(session,
 	    lsm_rows_merged, insert_count % LSM_MERGE_CHECK_INTERVAL);
+	++lsm_tree->merge_progressing;
 	WT_VERBOSE_ERR(session, lsm,
 	    "Bloom size for %" PRIu64 " has %" PRIu64 " items inserted.",
 	    record_count, insert_count);
diff --git a/src/lsm/lsm_tree.c b/src/lsm/lsm_tree.c
index 7f95ccff9de..135916afa3d 100644
--- a/src/lsm/lsm_tree.c
+++ b/src/lsm/lsm_tree.c
@@ -927,6 +927,52 @@ __wt_lsm_tree_unlock(
 }
 
 /*
+ * __wt_lsm_compact --
+ *	Compact an LSM tree called via __wt_schema_worker.
+ */
+int
+__wt_lsm_compact(WT_SESSION_IMPL *session, const char *name)
+{
+	WT_DECL_RET;
+	WT_LSM_TREE *lsm_tree;
+	uint64_t last_merge_progressing;
+	time_t begin, end;
+
+	/* Ignore non LSM names. */
+	if (!WT_PREFIX_MATCH(name, "lsm:"))
+		return (0);
+
+	WT_RET(__wt_lsm_tree_get(session, name, 0, &lsm_tree));
+
+	if (!F_ISSET(S2C(session), WT_CONN_LSM_MERGE) ||
+	    lsm_tree->merge_threads == 0)
+		WT_RET_MSG(session, EINVAL,
+		    "LSM compaction requires active merge threads");
+
+	WT_RET(__wt_seconds(session, &begin));
+
+	F_SET(lsm_tree, WT_LSM_TREE_COMPACTING);
+
+	/* Wake up the merge threads. */
+	WT_RET(__wt_cond_signal(session, lsm_tree->work_cond));
+
+	/* Now wait for merge activity to stop. */
+	do {
+		last_merge_progressing = lsm_tree->merge_progressing;
+		__wt_sleep(1, 0);
+		WT_RET(__wt_seconds(session, &end));
+		if (session->compact->max_time > 0 &&
+		    session->compact->max_time < (uint64_t)(end - begin))
+			WT_ERR(ETIMEDOUT);
+	} while (lsm_tree->merge_progressing != last_merge_progressing &&
+	    lsm_tree->nchunks > 1);
+
+err:	F_CLR(lsm_tree, WT_LSM_TREE_COMPACTING);
+
+	return (ret);
+}
+
+/*
  * __wt_lsm_tree_worker --
  *	Run a schema worker operation on each level of a LSM tree.
  */
diff --git a/src/lsm/lsm_worker.c b/src/lsm/lsm_worker.c
index d6c5011c33e..3cfb4119b85 100644
--- a/src/lsm/lsm_worker.c
+++ b/src/lsm/lsm_worker.c
@@ -445,7 +445,11 @@ __lsm_bloom_create(WT_SESSION_IMPL *session,
 		WT_RET(wt_session->drop(wt_session, chunk->bloom_uri, "force"));
 
 	bloom = NULL;
-
+	/*
+	 * This is merge-like activity, and we don't want compacts to give up
+	 * because we are creating a bunch of bloom filters before merging.
+	 */
+	++lsm_tree->merge_progressing;
 	WT_RET(__wt_bloom_create(session, chunk->bloom_uri,
 	    lsm_tree->bloom_config, chunk->count,
 	    lsm_tree->bloom_bit_count, lsm_tree->bloom_hash_count, &bloom));
diff --git a/src/schema/schema_worker.c b/src/schema/schema_worker.c
index 7c774e178d8..60143f97f77 100644
--- a/src/schema/schema_worker.c
+++ b/src/schema/schema_worker.c
@@ -52,9 +52,8 @@ __wt_schema_worker(WT_SESSION_IMPL *session,
 		WT_ERR(__wt_schema_worker(session, idx->source,
 		    file_func, name_func, cfg, open_flags));
 	} else if (WT_PREFIX_MATCH(uri, "lsm:")) {
-		if (file_func != __wt_compact)
-			WT_ERR(__wt_lsm_tree_worker(session,
-			    uri, file_func, name_func, cfg, open_flags));
+		WT_ERR(__wt_lsm_tree_worker(session,
+		    uri, file_func, name_func, cfg, open_flags));
 	} else if (WT_PREFIX_SKIP(tablename, "table:")) {
 		WT_ERR(__wt_schema_get_table(session,
 		    tablename, strlen(tablename), 0, &table));
diff --git a/src/session/session_api.c b/src/session/session_api.c
index 133348f2a01..67d43f39ff1 100644
--- a/src/session/session_api.c
+++ b/src/session/session_api.c
@@ -382,31 +382,19 @@ static int
 __session_compact(WT_SESSION *wt_session, const char *uri, const char *config)
 {
 	WT_SESSION_IMPL *session;
-	WT_TXN *txn;
 
 	session = (WT_SESSION_IMPL *)wt_session;
-	txn = &session->txn;
 
 	/* Disallow objects in the WiredTiger name space. */
 	WT_RET(__wt_schema_name_check(session, uri));
 
-	/* Compaction makes no sense for LSM objects, ignore requests. */
-	if (WT_PREFIX_MATCH(uri, "lsm:"))
-		return (0);
 	if (!WT_PREFIX_MATCH(uri, "colgroup:") &&
 	    !WT_PREFIX_MATCH(uri, "file:") &&
 	    !WT_PREFIX_MATCH(uri, "index:") &&
+	    !WT_PREFIX_MATCH(uri, "lsm:") &&
 	    !WT_PREFIX_MATCH(uri, "table:"))
 		return (__wt_bad_object_type(session, uri));
 
-	/*
-	 * Compaction requires checkpoints, which will fail in a transactional
-	 * context.  Check now so the error message isn't confusing.
-	 */
-	if (F_ISSET(txn, TXN_RUNNING))
-		WT_RET_MSG(session, EINVAL,
-		    "Compaction not permitted in a transaction");
-
 	return (__wt_session_compact(wt_session, uri, config));
 }
 
@@ -693,6 +681,7 @@ __session_checkpoint(WT_SESSION *wt_session, const char *config)
 	WT_TXN *txn;
 
 	session = (WT_SESSION_IMPL *)wt_session;
+
 	txn = &session->txn;
 
 	WT_STAT_FAST_CONN_INCR(session, txn_checkpoint);
diff --git a/src/session/session_compact.c b/src/session/session_compact.c
index 03d519004ec..d3f5be6b71b 100644
--- a/src/session/session_compact.c
+++ b/src/session/session_compact.c
@@ -96,63 +96,141 @@
  */
 
 /*
- * __session_compact_worker --
- *	Worker function to do the actual compaction call.
+ * __wt_compact_uri_analyze --
+ *	Extract information relevant to deciding what work compact needs to
+ *	do from a URI that is part of a table schema.
+ *	Called via the schema_worker function.
  */
-static int
-__session_compact_worker(
-    WT_SESSION *wt_session, const char *uri, const char *config)
+int
+__wt_compact_uri_analyze(WT_SESSION_IMPL *session, const char *uri)
 {
-	WT_DECL_RET;
-	WT_SESSION_IMPL *session;
+	/*
+	 * Add references to schema URI objects to the list of objects to be
+	 * compacted.
+	 */
+	if (WT_PREFIX_MATCH(uri, "lsm:"))
+		session->compact->lsm_count++;
+	else if (WT_PREFIX_MATCH(uri, "file:"))
+		session->compact->file_count++;
 
-	session = (WT_SESSION_IMPL *)wt_session;
-	SESSION_API_CALL(session, compact, config, cfg);
+	return (0);
+}
+
+/*
+ * __session_compact_check_timeout --
+ *	Check if the timeout has been exceeded.
+ */
+static int
+__session_compact_check_timeout(
+    WT_SESSION_IMPL *session, struct timespec begin)
+{
+	struct timespec end;
 
-	WT_WITH_SCHEMA_LOCK(session,
-	    ret = __wt_schema_worker(session, uri, __wt_compact, NULL, cfg, 0));
+	if (session->compact->max_time == 0)
+		return (0);
 
-err:	API_END_NOTFOUND_MAP(session, ret);
+	WT_RET(__wt_epoch(session, &end));
+	if (session->compact->max_time <
+	    WT_TIMEDIFF(end, begin) / WT_BILLION)
+		WT_RET(ETIMEDOUT);
+	return (0);
 }
 
 /*
- * __wt_session_compact --
- *	Function to alternate between checkpoints and compaction calls.
+ * __compact_file --
+ *
  */
-int
-__wt_session_compact(
-    WT_SESSION *wt_session, const char *uri, const char *config)
+static int
+__compact_file(WT_SESSION_IMPL *session, const char *uri, const char *cfg[])
 {
-	WT_DECL_ITEM(t);
 	WT_DECL_RET;
-	WT_SESSION_IMPL *session;
+	WT_DECL_ITEM(t);
+	WT_SESSION *wt_session;
+	WT_TXN *txn;
 	int i;
+	struct timespec start_time;
 
-	session = (WT_SESSION_IMPL *)wt_session;
+	txn = &session->txn;
+	wt_session = &session->iface;
+
+	/*
+	 * File compaction requires checkpoints, which will fail in a
+	 * transactional context.  Check now so the error message isn't
+	 * confusing.
+	 */
+	if (session->compact->file_count != 0 && F_ISSET(txn, TXN_RUNNING))
+		WT_ERR_MSG(session, EINVAL,
+		    " File compaction not permitted in a transaction");
 
 	/*
 	 * Force the checkpoint: we don't want to skip it because the work we
 	 * need to have done is done in the underlying block manager.
 	 */
-	WT_RET(__wt_scr_alloc(session, 128, &t));
+	WT_ERR(__wt_scr_alloc(session, 128, &t));
 	WT_ERR(__wt_buf_fmt(session, t, "target=(\"%s\"),force=1", uri));
 
+	WT_ERR(__wt_epoch(session, &start_time));
+
 	/*
 	 * We compact 10% of the file on each pass, try 10 times (which is
-	 * probably overkill), and quit if we make no progress.
+	 * probably overkill), and quit if we make no progress. Check for a
+	 * timeout each time through the loop.
 	 */
 	for (i = 0; i < 10; ++i) {
 		WT_ERR(wt_session->checkpoint(wt_session, t->data));
 
 		session->compaction = 0;
-		WT_ERR(__session_compact_worker(wt_session, uri, config));
+		WT_WITH_SCHEMA_LOCK(session,
+		    ret = __wt_schema_worker(
+		    session, uri, __wt_compact, NULL, cfg, 0));
+		WT_ERR(ret);
 		if (!session->compaction)
 			break;
 
 		WT_ERR(wt_session->checkpoint(wt_session, t->data));
 		WT_ERR(wt_session->checkpoint(wt_session, t->data));
+		WT_ERR(__session_compact_check_timeout(session, start_time));
 	}
 
 err:	__wt_scr_free(&t);
 	return (ret);
 }
+
+/*
+ * __wt_session_compact --
+ *	Function to alternate between checkpoints and compaction calls.
+ */
+int
+__wt_session_compact(
+    WT_SESSION *wt_session, const char *uri, const char *config)
+{
+	WT_COMPACT compact;
+	WT_CONFIG_ITEM cval;
+	WT_DECL_RET;
+	WT_SESSION_IMPL *session;
+
+	session = (WT_SESSION_IMPL *)wt_session;
+	SESSION_API_CALL(session, compact, config, cfg);
+
+	/* Setup the structure in the session handle */
+	memset(&compact, 0, sizeof(WT_COMPACT));
+	session->compact = &compact;
+
+	WT_ERR(__wt_config_gets(session, cfg, "timeout", &cval));
+	session->compact->max_time = (uint64_t)cval.val;
+
+	/* Find the types of data sources are being compacted. */
+	WT_WITH_SCHEMA_LOCK(session, ret = __wt_schema_worker(
+	    session, uri, NULL, __wt_compact_uri_analyze, cfg, 0));
+	WT_ERR(ret);
+
+	if (session->compact->lsm_count != 0)
+		WT_ERR(__wt_schema_worker(
+		    session, uri, NULL, __wt_lsm_compact, cfg, 0));
+	if (session->compact->file_count != 0)
+		WT_ERR(__compact_file(session, uri, cfg));
+
+err:	session->compact = NULL;
+	API_END_NOTFOUND_MAP(session, ret);
+	return (ret);
+}