Changes Gossiper to NodeSet.
Adds StaticNodeSet (for testing) and GossipNodeSet (for memberlist) implementations.
Changes NodeSet interface to return a pilosa-specific generic instead of `NodeSet interface`.
Removes NumMembers from interface (which is specific to memberlist).
Implements a Messenger interface with which to send inter-node messages via NodeSet.
The Pilosa implementation occurs in the GossipNodeSet.
Implements the Messenger as an object on Server, Handler, and Index.
Uses HealthStatus constants.
Removes commented-out code.
for gossip, make sure to bind to both host and port, and advertise those as well
adjust messenger to work with the db schema logic
add dependencies: hashicorp/memberlist, golang.org/x/sync
add dependency: golang.org/x/net
Uses `errgroup` to handle errors from broadcast messages.
Marshals message one time instead of once for every node.
Adds error handling for some errors that were being swallowed.
Previously the cache type (`lru` or `ranked`) was determined by
checking the ending suffix of the frame. If it ended with a `.n`
then it was ranked. Otherwise it was an LRU.
This commit changes frames so that a `cacheType` option can be
passed in during creation to set either `lru` or `ranked`.
Previously only the standard max slice was available. This commit
changes it so that the inverse max can be retrieved separately
through the `DB` and `Frame` types as well as through the
`HTTP` API and `Client`.
Previously, multiple frames with different prefixes were used to separate
different data layouts. This included separating standard row/column
layouts from inverted column/row layouts as well as storing aggregate
information for timestamp data.
Unfortunately, this caused frame meta data to be copied between multiple
frames and it made it difficult to keep these frames in sync.
This commit separates these different physical layouts into `Views`.
A `Frame` now has one or more views which represent each layout.
Fragments have been moved from under the `Frame` to be contained
within the `View`.
There are two primary views:
- `standard`
- `inverse`
If a frame has a time quantum, then views are generated for these
each of the standard/inverse views. For example a time quantum
of `YMDH` for the date `2000-01-02T00:00:00Z` would create the
following views:
- `standard_2000`
- `inverse_2000`
- `standard_200001`
- `inverse_200001`
- `standard_20000102`
- `inverse_20000102`
From the user's perspective, nothing should change in PQL. Different
PQL statements will handle the appropriate view automatically. For
example, `Bitmap()` and `Profile()` will fetch using the `standard`
view or the `inverse` view, respectively. The `Range()` statement
will lookup the appropriate time-based views automatically.
Databases and frames now require explicit creation and have the
option of setting row & column labels. If no labels are provided
then the default `id` and `profileID` labels are used.
- Ingore asserts in fragment container.
- Only log queries that take longer than 90 seconds.
TODO:
- address the TODOs that make the asserts configurable.
This changes the API of the types but does not alter the functionality
since only the `context.Background()` is currently being used. Adding
`Context` will help handle fault tolerance in the future by allowing
timeouts to be propagated across calls to different nodes.
CSV exports can now be done with the pilosactl application:
$ pilosactl -d mydb -f myframe -o MYFILE.csv
If `-o` is not specified then the CSV is written to STDOUT. The
exporter combines all slices for the db/frame to into a single
concatenated CSV file.
This commit changes the fragment allocation algorithm in the cluster
to make use of the `DB` name. This allows each database to use a
different slice distribution.
Initially, the `frame` was going to be used for allocation, however,
this was problematic since queries can span multiple frames so it's
impossible to choose a single frame to use.
This commit refactors the anti-entropy system to fetch data from
all replicated blocks and only set/clear bits which deviate from
the consensus between all blocks.
An example of this is if 3 nodes had the following bits set for
a single bitmap:
Node A: 1 2 3
Node B: 2 4
Node C: 1 2 4
Then only bits which are set on a majority will be set. In this
case bits 1, 2, & 4 are set but 3 only exists on a single node.
The node performing the merge would then determine the following
set/clear diffs for each node:
Node A: clear(3), set(4)
Node B: set(1)
Node C: none
Once the merge is performed and all nodes receive their diff
instructions then the nodes will be in sync:
Node A: 1 2 4
Node B: 1 2 4
Node C: 1 2 4
There still exists situations where bits can be reset. If Node A
is up and Node B & C are down then Node A's bits will be reset
once B & C come back online. We should add write consistency
settings for incoming writes so that we can ensure that a quorum
is written to before returning a success. This is outside the
scope of this commit though.
This commit adds a simple benchmarking utility to the `pilosactl`
binary. It currently only supports individual `SetBit()` commands
but it's a good start towards making a generic benchmarking
framework at the integration level.
The subcommands and usage/help messages were also cleaned up to
output correctly.
This commit adds a new HTTP handler to return a list of all databases
and frames in the index:
GET /schema
This returns JSON in the following format:
{
"dbs":[
{
"name":"d0",
"frames":[
{"name":"f0"},
{"name":"f1"}
]
}
]
}
Fixes#61
This commit adds an endpoint to restore an entire frame from
another cluster. Multiple hosts can use this endpoint to copy
and rebalance a cluster to a new cluster.
This commit changes the `pql.Query` so that it can accept one or more
top-level calls instead of only one.
The query request format change because a query with a single call is
still valid. However, the result format now returns a `results` field
that has one result for each top-level call. The `profiles` field is
still the same, however, it combines all profiles from all bitmap
responses into one return so that there's not duplicate attributes.
Fixes#59
This commit adds two commands:
pilosactl backup
pilosactl restore
The `Client` implementations have also been added to support these
commands:
func (c *Client) BackupTo(w io.Writer, db, frame string) error
func (c *Client) RestoreFrom(r io.Reader, db, frame string) error
Backups occur on a per-frame basis and all slices from the cluster
are packed into a single tar file. The backup tool attempts to
read from owner nodes in a random order and will retry against
the next owner if one fails.
During restore, the slices are restored to all owner nodes. This
allows users to backup a frame from one cluster and restore it
to a different one -- even if the topology of the new cluster is
different.
