Also, fix it so that heap inserts sort on tags as well.
This change makes it easier to implement bottom and also fixes the test that the previous change broke.
Instead of rounding up the points, sorting and then slicing, keep a
heap that allows us to quickly see if the point needs to be in the
set. This cuts a top query on a dataset of 8 million points from 35
seconds to 11 seconds.
Instead of using closures and 2 different sort routines, have an
interface compare method that makes it easy to switch directions
for comparisons.
Note that this changes the sort order of distinct to match that of
top. While it is a change, I don't think it will break any code. The
important thing for distinct is just that the ordering is absolute,
not what the order is.
Don't declare distinct stat map for partitions. It's more useful to see
the stats collated together per-WAL. This may need further change in the
future.
If the memory gets 5x above the partition size threshold, the WAL will start returning write failures to the clients. This will allow them to backoff their write volume.
Also updated the stress script to track failed requests and output messages on failure and when it returns to success.
With this change, the generic batcher used by many inputs can now be
buffered. Testing shows that this performance of the Graphite input by
10-100%, with the biggest improvements at lower numbers of connections.
* Only fire a go routine to flush and compact if it isn't already running
* Have a sleep backoff time that scales up as the percentage of memory used goes up
This commit fixes `SelectMapper.Open()` so that it properly rolls back
transactions. Previously, this caused transactions to stay open
indefinitely which caused mmap resizes to hang indefinitely.
Start of a lower-level file inspection tool. This currently dumps
summary statistics for the shards, index and WAL that can be used to
understand the shape of the data is in the local shards. This util
operates on the shards itself and not through the server and is intended
more for debugging/troubleshooting.
This change adds support for diagnostics by decomposing the existing
interface into two interfaces -- one for stats, and the other for
diags. It also adds some basic monitor of system, network, and the Go
runtime.
This commit converts meta.ShardInfo.OwnerIDs from a slice of ids
to a slice of objects. This is to support adding statuses for a
shard for a given node. For example, a node may have a shard
assigned to it but it is currently copying the shard and is not
ready to serve data for it.
The old `OwnerIDs` is marked as deprecated, however, the code
still supports loading from older protobuf-encoded data.
Running show measurements in a partially replicated cluster produces inconsistent
results due to the connection pooling. When running remote meta-data queries,
the cluster service ends ups keeping map shard request open but still checks the connection
back into the pool. This causes inconsistent results because data from the last request
interferes with the new request.
This removes the connection pool which fixes the issue. It also has the side effect of fixing
a nodes pool connections that have gone bad when a node restarts. For example, in a 3 node cluster
that has been responding to queries correctly, restarting 1 node will cause all the other to fail
to query that node indefinitely. This is now fixed as well.
Some nodes may receive requests for shards that they are assigned but
do not have the shards locally yet. Just return no results in this case
instead of panicing.
The TSDBStore was returning a nil mapper if the shard did not exist. The caller always
assumed the mapper would not be nil causing a panic. Instead, have the mapper skip the mapping
phase if it's shard reference is nil. This fixes queries against data-only nodes and against
shards that are not fully replicated in the cluster.
Fixes#3574
A write lock was being taken to read the memory size to determine if writes
should be paused. What happens is that writers get blocked indefintely when
trying to acquire a write lock which makes writes pause (or stop) for long periods
of time.
The log was deferring the release of the read lock on the WAL. This had
the affect that a read-lock was held until after the partition finished writing
(which maintains it's own locks). The read lock is only needed around the call
to pointsToPartions so it can get a consistent copy of the points to write. After
that calls returns, a lock is not needed so free it immediatedly.
addToCache is called in a goroutine and can panic if the server is closed while opening. If
part of the open func errors, it returns an error and immediately calls close. close sets
p.cache to nil which causes the goroutine trying to initialized the cache to panic as well. The
goroutine should run under a write lock to avoid this race/panic.
If LoadMetadataIndex() tries to log an error, it causes a panic because the
logger is not set until Open() is called, which is after LoadMetaDataIndex() returns.
Instead, just set the logger up when the WAL is created.
Writes could timeout and when adding new measurement names to the
index if the sort took a long time. The names slice was never
actually used (except a test) so keeping it in index wastes memory
and sort it wastes CPU and increases lock contention. The sorting
was happening while the shard held a write-lock.
Fixes#3869
If we've seeked a cursor, then we can be sure that there will be no
data between it and the point that was seeked to. Take advantage of
this fact to only seek when it would yield us a value that would be
different from the last.
In addition, only init the pointsheap when doing a raw query. For
aggregate queries, it is reinitialized on every time bucket, so no
need to seek through all the cursors
For a synthetic database where there was only entries for a tiny
slice of time, it cut queries from 112 seconds to 30 seconds doing
`select mean(value) from cpu where time > now - 2h group by time(1h)`
This commit changes the default block size from 64KB to 4KB for
bz1. This was lowered because small blocks were being uncompressed,
merged, recompressed, and inserted for a large portion of updates.
This became slower and slower over time until it reached the 64KB
threshold. We moved to the 4KB threshold in order to lower the
impact of this recompression.
Since we already got a pointsHeapItem, let's just reuse it instead
of allocating a new one. This cuts allocated memory of a 1 million
points aggregate query from 4881.97MB to 4139.86MB
The buffer allocation in bz1 was unused and I'm fairly certain that it
was harmful to performance if used. For queries that run through a bz1
block, needing to hold on to a 64kb block is expensive. Better to churn
on the allocator and have the blocks be released when they are unused
than to have 64kb hanging around for each series regardless of size.
Thanks to @jwilder for brainstorming this issue with me.
By using preallocated buffers for marshaling WAL entries, we can
reduce the amount of memory we allocate.
On a run of `influx_stress -series 10000 -points 1000` this cuts
total allocations from 18684.15MB to 15200.73MB
* Update the store to remove the WAL directories associated with a shard or database when they are deleted.
* Fix the Store so that it creates separate WAL directories for databases and retention policies.
This func show up in profiling. It's called frequently from multiple places and
can be made more efficient. The previous implementation looped over the input
slice 4 times updating an returning a new slice each time. The changes it to loop
once and create one result slice.
With influx_stress
Before:
Wrote 10000000 points at average rate of 241750
Average response time: 187.78968ms
After:
Wrote 10000000 points at average rate of 254618
Average response time: 172.235028ms
Through profiling of writes, point.Fields() and point.Name() were called
repeatedly in PointsWriter and the Shard. These calls are somewhat expensive
when writing large batches so we can cache them to avoid wasting CPU cycles.
Using influx_stress with default settings
Before:
Wrote 10000000 points at average rate of 202570
Average response time: 235.450355ms
After:
Wrote 10000000 points at average rate of 246120
Average response time: 182.881008ms
This commit changes the bz1 append to check for a small
ending block first. If the block is below the threshold
for block size then it is rewritten with the new data
points instead of having a new block written.
If a flush is happening and you bring up a cursor for a series, if that series didn't have any data in the cache (after the flush started) then it would return no data. What it should have done instead is return the data that is in the flush cache, which is held in separate area of memory until it is committed to the index.
If the measurement started with a quote, a panic would happen. This
is a reegression due to cb7f0b8.
This also uncovered that measurement names were being escaped incorrectly.
The escape codes for tag and fields also includes `=` and '"` which should
not be escaped for measurement names.
Fixes#3681
* All metadata for each shard is now stored in a single key with compressed value
* Creation of new metadata no longer requires a syncrhnous write to Bolt. It is passed to the WAL and written to Bolt periodically outside the write path
* Added DeleteSeries to WAL and updated bz1 to remove series there when DeleteSeries or DropMeasurement are called
CPU profiling shows that computing the tagset-based key of each point is significant CPU cost. These keys actually don't change per cursor, so precompute once at mapper-open time, and then use those values as points are drained from the cursor.
Before this change the cursor tag was getting computed on every point, which involved marshalling tags.
The series map on Measurement was updated and deleted from but never
actually used. Series keys can be very bia since they are the the
string representation of the measurement plus sorted tags.
Locally I see 20%-30% reduction in memory usage with 1M series.
dgnorton
Ben Johnson
Nathaniel Cook
Philip O'Toole
Nathaniel Cook
Philip O'Toole
Daniel Morsing
Mark Rushakoff
Daniel Morsing
Cory LaNou
Cameron Sparr
Paul Dix
Jason Wilder
Todd Persen
Paul Dix