Prometheus needs long-term storage. Since we don't have enough resources
to build our own timeseries storage from scratch ontop of Riak,
Cassandra or a similar distributed datastore at the moment, we're
planning on using OpenTSDB as long-term storage for Prometheus. It's
data model is roughly compatible with that of Prometheus, with some
caveats.
As a first step, this adds write-only replication from Prometheus to
OpenTSDB, with the following things worth noting:
1)
I tried to keep the integration lightweight, meaning that anything
related to OpenTSDB is isolated to its own package and only main knows
about it (essentially it tees all samples to both the existing storage
and TSDB). It's not touching the existing TieredStorage at all to avoid
more complexity in that area. This might change in the future,
especially if we decide to implement a read path for OpenTSDB through
Prometheus as well.
2)
Backpressure while sending to OpenTSDB is handled by simply dropping
samples on the floor when the in-memory queue of samples destined for
OpenTSDB runs full. Prometheus also only attempts to send samples once,
rather than implementing a complex retry algorithm. Thus, replication to
OpenTSDB is best-effort for now. If needed, this may be extended in the
future.
3)
Samples are sent in batches of limited size to OpenTSDB. The optimal
batch size, timeout parameters, etc. may need to be adjusted in the
future.
4)
OpenTSDB has different rules for legal characters in tag (label) values.
While Prometheus allows any characters in label values, OpenTSDB limits
them to a to z, A to Z, 0 to 9, -, _, . and /. Currently any illegal
characters in Prometheus label values are simply replaced by an
underscore. Especially when integrating OpenTSDB with the read path in
Prometheus, we'll need to reconsider this: either we'll need to
introduce the same limitations for Prometheus labels or escape/encode
illegal characters in OpenTSDB in such a way that they are fully
decodable again when reading through Prometheus, so that corresponding
timeseries in both systems match in their labelsets.
Change-Id: I8394c9c55dbac3946a0fa497f566d5e6e2d600b5
So far we've been using Go's native time.Time for anything related to sample
timestamps. Since the range of time.Time is much bigger than what we need, this
has created two problems:
- there could be time.Time values which were out of the range/precision of the
time type that we persist to disk, therefore causing incorrectly ordered keys.
One bug caused by this was:
https://github.com/prometheus/prometheus/issues/367
It would be good to use a timestamp type that's more closely aligned with
what the underlying storage supports.
- sizeof(time.Time) is 192, while Prometheus should be ok with a single 64-bit
Unix timestamp (possibly even a 32-bit one). Since we store samples in large
numbers, this seriously affects memory usage. Furthermore, copying/working
with the data will be faster if it's smaller.
*MEMORY USAGE RESULTS*
Initial memory usage comparisons for a running Prometheus with 1 timeseries and
100,000 samples show roughly a 13% decrease in total (VIRT) memory usage. In my
tests, this advantage for some reason decreased a bit the more samples the
timeseries had (to 5-7% for millions of samples). This I can't fully explain,
but perhaps garbage collection issues were involved.
*WHEN TO USE THE NEW TIMESTAMP TYPE*
The new clientmodel.Timestamp type should be used whenever time
calculations are either directly or indirectly related to sample
timestamps.
For example:
- the timestamp of a sample itself
- all kinds of watermarks
- anything that may become or is compared to a sample timestamp (like the timestamp
passed into Target.Scrape()).
When to still use time.Time:
- for measuring durations/times not related to sample timestamps, like duration
telemetry exporting, timers that indicate how frequently to execute some
action, etc.
*NOTE ON OPERATOR OPTIMIZATION TESTS*
We don't use operator optimization code anymore, but it still lives in
the code as dead code. It still has tests, but I couldn't get all of them to
pass with the new timestamp format. I commented out the failing cases for now,
but we should probably remove the dead code soon. I just didn't want to do that
in the same change as this.
Change-Id: I821787414b0debe85c9fffaeb57abd453727af0f
This fixes part 2) of https://github.com/prometheus/prometheus/issues/367
(uninitialized time.Time mapping to a higher LevelDB key than "normal"
timestamps).
Change-Id: Ib079974110a7b7c4757948f81fc47d3d29ae43c9
This fixes part 1) of https://github.com/prometheus/prometheus/issues/367 (the
storing of samples with the wrong fingerprint into a compacted chunk, thus
corrupting it).
Change-Id: I4c36d0d2e508e37a0aba90b8ca2ecc78ee03e3f1
This commit fixes a critique of the old storage API design, whereby
the input parameters were always as raw bytes and never Protocol
Buffer messages that encapsulated the data, meaning every place a
read or mutation was conducted needed to manually perform said
translations on its own. This is taxing.
Change-Id: I4786938d0d207cefb7782bd2bd96a517eead186f
While a hack, this change should allow us to serve queries
expeditiously during a flush operation.
Change-Id: I9a483fd1dd2b0638ab24ace960df08773c4a5079
The background curation should be staggered to ensure that disk
I/O yields to user-interactive operations in a timely manner. The
lack of routine prioritization necessitates this.
Change-Id: I9b498a74ccd933ffb856e06fedc167430e521d86
Move the stream to an interface, for a number of additional changes
around it are underway.
Conflicts:
storage/metric/memory.go
Change-Id: I4a5fc176f4a5274a64ebdb1cad52600954c463c3
AppendSample will be repcated with AppendSamples, which will take
advantage of bulks appends. This is a necessary step for indexing
pipeline decoupling.
Change-Id: Ia83811a87bcc89973d3b64d64b85a28710253ebc
This commit is the first of several and should not be regarded as the
desired end state for these cleanups. What this one does it, however,
is wrap the query index writing behind an interface type that can be
injected into the storage stack and have its lifecycle managed
separately as needed. It also would mean we can swap out underlying
implementations to support remote indexing, buffering, no-op indexing
very easily.
In the future, most of the individual index interface members in the
tiered storage will go away in favor of agents that can query and
resolve what they need from the datastore without the user knowing
how and why they work.
There are too many parameters to constructing a LevelDB storage
instance for a construction method, so I've opted to take an
idiomatic approach of embedding them in a struct for easier
mediation and versioning.
When samples get flushed to disk, they lose sub-second precision anyways. By
already dropping sub-second precision, data fetched from memory vs. disk will
behave the same. Later, we should consider also storing a more compact
representation than time.Time in memory if we're not going to use its full
precision.
Current series always get watermarks written out upon append now. This
drops support for old series without any watermarks by always reporting
them as too old (stale) during queries.
This also short-circuits optimize() for now, since it is complex to implement
for the new operator, and ops generated by the query layer already fulfill the
needed invariants. We should still investigate later whether to completely
delete operator optimization code or extend it to support
getValueRangeAtIntervalOp operators.
An design question was open for me in the beginning was whether to
serialize other types to disk, but Protocol Buffers quickly won out,
which allows us to drop support for other types. This is a good
start to cleaning up a lot of cruft in the storage stack and
can let us eventually decouple the various moving parts into
separate subsystems for easier reasoning.
This commit is not strictly required, but it is a start to making
the rest a lot more enjoyable to interact with.
This adds timers around several query-relevant code blocks. For now, the
query timer stats are only logged for queries initiated through the UI.
In other cases (rule evaluations), the stats are simply thrown away.
My hope is that this helps us understand where queries spend time,
especially in cases where they sometimes hang for unusual amounts of
time.
This commit conditionalizes the creation of the diskFrontier and
seriesFrontier along with the iterator such that they are provisioned
once something is actually required from disk.
This is mainly a small performance improvement, since we skip past the last
extracted time immediately if it was also the last sample in the chunk, instead
of trying to extract non-existent values before the chunk end again and again
and only gradually approaching the end of the chunk.
The current behavior only adds those samples to the view that are extracted by
the last pass of the last processed op and throws other ones away. This is a
bug. We need to append all samples that are extracted by each op pass.
This also makes view.appendSamples() take an array of samples.
The previous implementation spawned N goroutines to group samples
together and would not start work until the semaphore unblocked.
While this didn't leak, it polluted the scheduling space. Thusly,
the routine only starts after a semaphore has been acquired.
The one-off keys have been replaced with ``model.LabelPair``, which is
indexable. The performance impact is negligible, but it represents
a cognitive simplification.
The reality is that if we ever try to encode a Protocol Buffer and it
fails, it's likely that such an error is ultimately not a runtime error
and should be fixed forthwith. Thusly, we should rename
``Encoder.Encode`` to ``Encoder.MustEncode`` and drop the error return
value.
Some users of GetMetricForFingerprint() end up modifying the returned metric
labelset. Since the memory storage's implementation of
GetMetricForFingerprint() returned a pointer to the metric (and maps are
reference types anyways), the external mutation propagated back into the memory
storage.
The fix is to make a copy of the metric before returning it.
- only the data extracted in the last loop iteration of ExtractSamples() was
emitted as output
- if e.g. op interval < sample interval, there were situations where the same
sample was added multiple times to the output
This commit updates the documentation, Makefiles, formatting, and
code semantics to support the 1.1. runtime, which includes ...
1. ``make advice``,
2. ``make format``, and
3. ``go fix`` on various targets.
This commit simplifies the way that compactions across a database's
keyspace occur due to reading the LevelDB internals. Secondarily it
introduces the database size estimation mechanisms.
Include database health and help interfaces.
Add database statistics; remove status goroutines.
This commit kills the use of Go routines to expose status throughout
the web components of Prometheus. It also dumps raw LevelDB status
on a separate /databases endpoint.
This commit simplifies the way that compactions across a database's
keyspace occur due to reading the LevelDB internals. Secondarily it
introduces the database size estimation mechanisms.
This commit introduces the long-tail deletion mechanism, which will
automatically cull old sample values. It is an acceptable
hold-over until we get a resampling pipeline implemented.
Kill legacy OS X documentation, too.
This does two things:
1) Make TieredStorage.AppendSamples() write directly to memory instead of
buffering to a channel first. This is needed in cases where a rule might
immediately need the data generated by a previous rule.
2) Replace the single storage mutex by two new ones:
- memoryMutex - needs to be locked at any time that two concurrent
goroutines could be accessing (via read or write) the
TieredStorage memoryArena.
- memoryDeleteMutex - used to prevent any deletion of samples from
memoryArena as long as renderView is running and
assembling data from it.
The LevelDB disk storage does not need to be protected by a mutex when
rendering a view since renderView works off a LevelDB snapshot.
The rationale against adding memoryMutex directly to the memory storage: taking
a mutex does come with a small inherent time cost, and taking it is only
required in few places. In fact, no locking is required for the memory storage
instance which is part of a view (and not the TieredStorage).
This commit extracts the model.Values truncation behavior into the actual
tiered storage, which uses it and behaves in a peculiar way—notably the
retention of previous elements if the chunk were to ever go empty. This is
done to enable interpolation between sparse sample values in the evaluation
cycle. Nothing necessarily new here—just an extraction.
Now, the model.Values TruncateBefore functionality would do what a user
would expect without any surprises, which is required for the
DeletionProcessor, which may decide to split a large chunk in two if it
determines that the chunk contains the cut-off time.
This commit introduces three background compactors, which compact
sparse samples together.
1. Older than five minutes is grouped together into chunks of 50 every 30
minutes.
2. Older than 60 minutes is grouped together into chunks of 250 every 50
minutes.
3. Older than one day is grouped together into chunks of 5000 every 70
minutes.
This commit drops the Storage interface and just replaces it with a
publicized TieredStorage type. Storage had been anticipated to be
used as a wrapper for testability but just was not used due to
practicality. Merely overengineered. My bad. Anyway, we will
eventually instantiate the TieredStorage dependencies in main.go and
pass them in for more intelligent lifecycle management.
These changes will pave the way for managing the curators without
Law of Demeter violations.
This commit employs explicit memory freeing for the in-memory storage
arenas. Secondarily, we take advantage of smaller channel buffer sizes
in the test.
The curator requires the existence of a curator remark table, which
stores the progress for a given curation policy. The tests for the
curator create an ad hoc table, but core Prometheus presently lacks
said table, which this commit adds.
Secondarily, the error handling for the LevelDB lifecycle functions
in the metric persistence have been wrapped into an UncertaintyGroup,
which mirrors some of the functions of sync.WaitGroup but adds error
capturing capability to the mix.
This commit introduces to Prometheus a batch database sample curator,
which corroborates the high watermarks for sample series against the
curation watermark table to see whether a curator of a given type
needs to be run.
The curator is an abstract executor, which runs various curation
strategies across the database. It remarks the progress for each
type of curation processor that runs for a given sample series.
A curation procesor is responsible for effectuating the underlying
batch changes that are request. In this commit, we introduce the
CompactionProcessor, which takes several bits of runtime metadata and
combine sparse sample entries in the database together to form larger
groups. For instance, for a given series it would be possible to
have the curator effectuate the following grouping:
- Samples Older than Two Weeks: Grouped into Bunches of 10000
- Samples Older than One Week: Grouped into Bunches of 1000
- Samples Older than One Day: Grouped into Bunches of 100
- Samples Older than One Hour: Grouped into Bunches of 10
The benefits hereof of such a compaction are 1. a smaller search
space in the database keyspace, 2. better employment of compression
for repetious values, and 3. reduced seek times.
For the forthcoming Curator, we don't record timezone information in
the samples, nor do we in the curation remarks. All times are
recorded UTC. That said, for the test environment to better match
production, the special instant should be in UTC.
The curator work can be done easier if dto.SampleKey is no longer
directly accessed but rather has a higher level type around it that
captures a certain modicum of business logic. This doesn't look
terribly interesting today, but it will get more so.
After SampleValue was refactored into SampleValueSeries, which
involves plural values under a common super key, the stochastic
test was never refreshed to reflect this reality. We had other
tests that validated the functionality, but this one was
insufficently forward-ported.
This makes the memory persistence the backing store for views and
adjusts the MetricPersistence interface accordingly. It also removes
unused Get* method implementations from the LevelDB persistence so they
don't need to be adapted to the new interface. In the future, we should
rethink these interfaces.
All staleness and interpolation handling is now removed from the storage
layer and will be handled only by the query layer in the future.
The original append queue telemetry never worked, because it was
updated only upon the exit of the select statement, which would
usually liberate the queues of contents. This has been fixed to
be reported arbitrarily.
The queue sizes are now parameterizable via flags.
It is the case with the benchmark tool that we thought that we
generated multiple series and saved them to the disk as such, when
in reality, we overwrote the fields of the outgoing metrics via
Go map reference behavior. This was accidental. In the course of
diagnosing this, a few errors were found:
1. ``newSeriesFrontier`` should check to see if the candidate fingerprint is within the given domain of the ``diskFrontier``. If not, as the contract in the docstring stipulates, a ``nil`` ``seriesFrontier`` should be emitted.
2. In the interests of aiding debugging, the raw LevelDB ``levigoIterator`` type now includes a helpful forensics ``String()`` method.
This work produced additional cleanups:
1. ``Close() error`` with the storage stack is technically incorrect, since nowhere in the bowels of it does an error actually occur. The interface has been simplified to remove this for now.
After this commit, we'll need to add validations that it does the
desired work, which we presently know that it doesn't. Given the
changes I made with a plethora of renamings, I want to commit this
now before it gets even larger.
The curator doesn't do anything yet; rather, this is the type
definition including the anciliary testing scaffold.
Improve Makefile and Git developer experience.
The top-level Makefile was a bit overloaded in terms of generation of
assets and their management. This has been offloaded into separate
Makefiles.
The Git developer experience sucked due to lack of .gitignore
policies.
Also: Fix faulty skiplist naming from old merge.
The LevelDB storage types return an interface type now that wraps
around the underlying iterator. This both enhances testability but
improves upon, in my opinion, the interface design for the LevelDB
iterator.
Secondarily, the resource reaping behaviors for the LevelDB iterators
have been improved by dropping the externalized io.Closer object.
Finally, the iterator provisioning methods provide the option for
indicating whether one wants a snapshotted iterator or not.
EachFunc is deprecated.
Remove deprecated ``Pair`` and ``GetAll``.
These were originally used for forensic and the old gorest impl.
Nothing today in the user-facing path nor the tests uses them,
especially since the advent of the ForEach protocol in the
interface.
- Kill Close in Persistent and document interface.
- Extract batching behavior into interface.
- Kill IteratorManager, which was used for unknown reasons.
The LevelDB index retrievals could be repeated in a given operation
batch if multiple queued mutations affect the same (Label Name) singles
and (Label Name, Label Value) doubles. This is wasteful and
inefficient, as a single retrieval suffices. Thusly this commit
retrieves the canonical index mappings if the said mapping has not
been looked up in a given batch.
Julius Volz